Skip to content

AgentRun v0.1.0-beta.38

Pre-release
Pre-release

Choose a tag to compare

@claudio-michel claudio-michel released this 14 Nov 23:10
· 513 commits to main since this release
d10f005

AgentRun 0.1.0-beta.38

AgentRun is an open-source platform that supercharges your agents. It adds self-improvement, durable execution and a
control plane on top of your favorite agent harness. Your agents become smarter, more fault tolerant and use fewer
tokens with every run. This release is about tools, templates and what an agent learns. Every harness (pi, Claude Code,
Codex, OpenCode) now calls MCP tools through one gate, finds tools with tools_search and runs scripts with codemode.
A template package installs from a GitHub address, names where its tools come from, takes typed input and answers in a
schema. A learning agent compiles its job into a learned workflow, runs it first, and serves a new version only after a
replay on graded runs. A Linux desktop opens apps and reads plainly in the timeline, and a program inside a computer can
no longer reach the desktop's session bus. A capped Codex, OpenCode or pi run pauses less often. There is no fresh store
this time: the SQLite store stays at version 11 and Postgres migrates itself. The launcher is @parcha/agentrun 0.4.6,
the Python client parcha-agentrun is 0.3.6, and the server image is
ghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.38 (linux/amd64 and linux/arm64).

No fresh store: beta.38 opens the data beta.37 left. The SQLite schema is still version 11, and Postgres goes from
9 to 11 (migrations 0010 and 0011). Keep the same data folder and the same start command. On the Docker image that is
--data /data/b37 --listen 0.0.0.0 --port 7340 --drain-timeout 6s as before, and on Render the Docker Command you set
for beta.37 stays. Change only the image to ghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.38. The console keeps its
login, and your agents, MCP servers, keys and sign-ins are all still there.

  • SQLite. Two changes, both made by the server when it opens your folder, so there is nothing to run: a new table,
    imported_templates (templates installed from GitHub, #2728), and an agent revision's MCP pin in a new column,
    revisions.mcp_pin, in place of the per-run pin table mcp_pins, which is dropped (#2719). Copy the data folder
    before you upgrade.
    beta.37 still opens the folder afterwards (it checks the schema version only, 11), but it writes
    each run's MCP pin to mcp_pins at the run's first claim, so a run that grants an MCP tool fails on beta.37 once
    beta.38 has dropped that table. To go back, restore the copy and set the image back to v0.1.0-beta.37; the command
    stays.
  • Postgres. Migrations 0010_imported_templates.sql and 0011_revision_mcp_pin.sql run when the server starts,
    and the server moves the schema to version 11. Going back does not work without a backup: beta.37 refuses a
    schema at version 11, with the Postgres schema agentrun is schema version 11, set up by a later release than this one (this release knows version 9). Run that release or a later one. Take a backup of the store's schema before
    you upgrade
    (for example pg_dump --schema=agentrun <url>, with your own schema name if you set one), and to go
    back restore it and run beta.37 on the restored data. Start the server first: a beta.38 runner started on a schema
    the server has not yet migrated refuses to start, and so does a beta.37 runner on a migrated one, so upgrade the
    server, then the runners.
  • Nothing else under the store changed shape. Agents you saved on beta.38 may carry fields beta.37 does not know
    (input_schema, a computer's surface and exec, a template record); beta.37 reads past them.

Highlights

  • One tool plane on every harness.
    • A run's MCP tools are host tools on every harness: each call passes the run's one tool gate, the runner calls the
      server with the credential and books it, and OpenCode now runs MCP tools too (#2722).
    • Every harness offers tools_search (find a tool by what it does), tools_call (run one by name) and codemode (a
      JavaScript script over the run's tools), each gated, recorded and booked as the tool's own call (#2749). Past 20
      catalog tools (MCP and monid) a run defers them behind search, and each of the three has a setting you can deny
      (#2759).
    • On pi, a script's or tools_call's nested calls show in the timeline as on the other harnesses (#2767).
    • One grammar for tool names, one MCP scrubber, cap and client, and grant_problems served on a saved server say why
      a tool can't be granted (#2690, #2738).
  • Templates from GitHub. Install a template package from github:<owner>/<repo>[/<path>], pinned to a commit, with
    an "Add from URL" box on Templates, agentrun template install on the CLI and a page that shows the publisher, the
    commit and every file before anything is stored (#2728). A template's MCP needs now name where their tools come from
    (Composio, GitHub or a URL), and apply maps each need to the saved server that matches (#2732). Apply grants only the
    tools you tick (#2718). The install and Apply screens show each run input, eval case and tool reason (#2789).
  • Typed run input. An agentic agent can take a JSON Schema input (definition.input_schema); every path that starts
    a run checks it, and the console draws the run form from it (#2731). A template parameter is an input in the same
    grammar (#2775), and a list parameter fills a schedule's typed input (#2733).
  • An answer that meets its schema. An agent with an output_schema is told the schema on every harness, is asked
    once more when its answer misses it, and the run serves output_error if it still does; the console's Report reads
    the structured answer (#2725).
  • Learning that compiles and proves itself.
    • A learning agent writes its whole job as a learned workflow in a compile turn, with a record of each attempt in
      GET /runs/{id}/learning (#2641).
    • A run starts on a fitting learned workflow before the agent's own model takes a turn, and a money rule asks for a
      new version when the workflow costs more than the agent by hand (#2603).
    • A workflow step can carry a guard that hands the job back below a threshold (#2639).
    • A new version serves first only after a replay on the agent's graded runs, with thresholds refit from them (#2644).
    • The console's Learning tab walks through the loop: a five-stage strip says what is learned and what happens next, and
      the cost chart flags each new version, promotion, hold and refit on the run it first affected (#2654).
  • A Linux desktop that opens apps (#2727, #2735). computer.apps lists, opens and closes apps on E2B and Agent37
    desktops, with a window manager so they are framed and focused, and a Mac and a Linux desktop read the same way.
  • A run on a computer reads plainly. The timeline says what the agent did ("Pressed “Display”", "Typed “dark”"), one
    voice in the row, the call list and a Mac movie, with a code-mode script's calls on its own row (#2729, #2762,
    #2747). The Screen tab reads sooner and has one Take over (#2730, #2745, #2736).
  • Computer isolation. A program running inside an E2B or Agent37 computer can no longer reach the desktop's
    accessibility bus or session bus (#2780, #2787, #2790, with the check in #2723). Nothing about what a computer can do
    for the agent changes: nsenter and rootless podman still work inside a computer.
  • Caps, tighter. A capped Codex, OpenCode or pi (on OpenAI's Responses) run's next-step estimate is closer to its
    real input, so it pauses for a raise later and less often (#2772). The wrap-up note, the reserve and admission price a
    warm request alike, so a capped Codex or Claude Code run is no longer told to finish with most of its cap left
    (#2786).
  • ask_person on Claude Code and Codex (#2743): a Claude Code or Codex agent can ask you a question mid-run. The run
    parks waiting for your reply, and your answer resumes it.

Before you upgrade

  • A gateway must serve /v1/messages for Claude (#2799). With AGENT_MODEL_BASE_URL set, pi now sends a Claude model
    to <base>/v1/messages (as LiteLLM's proxy and OpenRouter serve it) instead of /v1/responses. A gateway that serves
    only Responses fails those runs. If you set AGENT_MODEL_PRICE for a Claude model, give it the fourth number, the
    cache-write price (Anthropic's is 1.25 times input): with three numbers, a cache write is booked at the input price.
  • A pinned tool takes only its listed arguments (#2798). A call of an MCP tool that has a pin, in the agent's settings
    or the server's, is refused when it carries a top-level argument the tool's listing does not declare under properties
    (and is not pinned), and the model reads which one. A pin that relied on a server's undeclared alias (Slack's
    channel_id, Gmail's to) now fails that call: use the listed name, or drop the pin. Tools with no pin are unchanged.
  • A label goes with the grade (#2648): a client that sent a grader's label as a follow-up message
    (POST /runs/{id}/messages) sends it as the grade's label (PUT /runs/{id}/grade) instead.
  • monid is granted per endpoint (#2765). monid.discover, monid.inspect and monid.run are gone: an agent that
    grants one is refused at its save (422 not_an_option). Grant the endpoints it calls instead, monid:<provider>/<endpoint>
    (say monid:apollo/people/match) or monid:<provider>/* (at most 50), with a monid key saved in Settings → Tool keys.
    Each is pinned at the save with its description, input schema and price, and is one tool of its own on every
    harness. An endpoint monid prices by usage can't be granted (no hold could keep the cap), and a workflow can't bind a
    monid endpoint yet.
  • An agent's MCP tools are pinned when it is saved (#2719), on its revision, not at each run's first claim. Every run
    of that revision calls the tools as the server listed them at the save, with their prices.
    • A save (a create, a definition PUT, a template's apply, an inline agent's first run of a version) is refused with 422
      on the tool's field when a granted server can't be listed right now or no longer lists a granted tool, and nothing
      is saved: save again once the server answers. Applying a template with a GitHub tool ticked on a server with no
      GitHub App connected is refused this way: connect GitHub in Settings and apply again.
    • To pick up a tool a server changed, save the agent again: that is a new revision with the new pin.
    • A revision saved before this release pins at its first run, once; nothing needs saving again. If one of its servers
      can't be listed then, the run is refused (409 agent_not_runnable) and the next run tries again.
  • The store. No fresh store; see the note at the top. Copy the SQLite data folder, or back up the Postgres schema,
    first (#2719, and migration 0010 on Postgres): going back to beta.37 needs that copy.
  • Tool-call budgets are gone (#2771). max_tool_calls on an agent definition and budget.tool_calls in a template
    manifest are refused as unknown fields, and the subagent tool takes no max_tool_calls. The count only ever held on
    pi; the run's spend cap and its wrap-up stop a runaway loop on every harness. Remove the field, put a count in the
    instructions if you want one, and set per_run_usd (a template's budget is now {per_run_usd}, required when
    budget is given).
  • MCP tools run as host tools (#2722). A run parked on an MCP call before the upgrade does not resume that call:
    cancel it and start it again. A Codex thread started before the upgrade loses its MCP tools when it resumes: start a
    new run. claude-agent offers a granted tool as mcp__hands__mcp__<server>__<tool> and codex as mcp_<server>__<tool>;
    the Timeline, tool settings and approvals still name it mcp:<server>/<tool>.
  • Tool names are held to one rule (#2690): mcp:<server>/<tool>, the server's name in lowercase letters, digits and
    - (at most 32, starting with a letter), the tool's in letters, digits, _ and - (at most 64). A tool setting or
    template that names a tool with a ., a capital in the server name or an over-long name now gets 422 not_allowed.
    Rename it to the server's listed name, or drop it: no run could call it.
  • Tools you have not registered ask by default (#2625): mcp:<server>/<tool> and a monid endpoint (#2765) default
    to ask, not allow. No tool of today's table changes.
  • Capped runs carry three more tool definitions (#2749): tools_search, tools_call and codemode ride on every
    model request, about 2,000 bytes (639 tokens on Sonnet), so the smallest caps rise. A pi desktop run on Sonnet through
    a gateway needs $0.07 for its first request and $0.13 for five steps in strict mode, or $0.07 in expected mode (it
    was $0.06, $0.12 and $0.06). Straight to Anthropic it needs $0.04 for its first request. A run you cap at the old
    floor pauses for a raise.
  • Templates.
    • Parameters (#2775): a parameter is required unless it says required: false, and a default is allowed only
      beside required: false. A package that left a parameter optional by omitting required adds required: false.
      GET /v2/templates serves parameters as a closed JSON Schema object, so a client that read
      parameters.<name>.required reads parameters.required and parameters.properties.<name>.
    • MCP needs (#2732): server: <name> in a template's mcp is removed; a package that still has it is refused with
      the new shape in the message. Replace it with source (composio, github or url) and toolkit. Inside the
      package its tools are mcp:<toolkit>/<tool>, and an apply's grants name them that way.
    • code-review and release-notes (#2732): their repo parameter is now owner and repo, their GitHub tools are
      pinned to that one repository, and the free-text searches that could not be pinned are dropped. Apply them again to
      get the new shape; an agent applied earlier keeps its own.
    • inbound-lead (#2732) reads nothing from HubSpot; its one HubSpot tool is an optional note on the contact the
      event names.
    • market-monitor and inbound-lead (#2810) no longer pin Slack's channel_id or Gmail's to: the gate refuses an
      argument a pinned tool doesn't list (#2798).
    • Unticked tools are not granted (#2718): an agent applied from a template before this release may still hold every
      tool of its MCP servers. Review its tools (or apply the template again) and remove the ones you did not mean to grant,
      above all on an agent set to never.
    • Package folders (#2774): the schema grader is gone (a graded run of a package that declares output is held to
      its schema by an output verdict, and a case with graders/schema.md is refused), the examples/ folder and the
      checks field in agent.md are refused, and GET /v2/templates no longer serves examples.
    • kyb (#2731, #2766): it takes its subject as typed input (company, optionally website, country,
      registration_number, reason); a newly applied kyb refuses a run without company. Its report's subject.input
      holds those names, where it held name and domain. An agent applied from the earlier kyb keeps its own definition.
    • Answer schema (#2725): applying a template no longer copies its answer schema beside a skill's SKILL.md, because
      the runner gives the agent the schema itself. code-review's findings.schema.json is now
      skills/review/findings.schema.json.
  • Agents with an output schema (#2725) have the schema appended to their instructions and get one more turn when the
    answer misses it, booked like any turn, under the run's cap.
  • Learning (#2644, #2603): a learned workflow draft that gives two steps one label is refused at learn time, as a
    saved one already was. A learning agent whose workflow fits a run's input now answers it before the agent's model is
    asked; set workflow_first: false on a run to send it to the agent from its first turn.
  • Computers.
    • Daytona (#2624, #2742): a Daytona computer's network can no longer be changed after create (it is refused), and
      Daytona takes no Archil mount: a create that asks for one is refused up front, and DAYTONA_ARCHIL_EGRESS is no
      longer read. A Mode 2 agent on a Daytona sandbox is refused at its save and its run's start. A Daytona computer made before this release carries no agentrun.computer.name label, so the provider no
      longer lists, attaches to or releases it: delete it in Daytona.
    • Archil label (#2712): an Archil computer made before this release on E2B, Daytona or Agent37 carries the mount
      in the older label form, so the provider reads it as one it never makes: delete it at the vendor. Modal's already
      held the new form.
    • Daemon version (#2715): a computer's daemon must match its provider's version and protocol exactly; a daemon
      whose hello names no protocol is refused.
    • Agent37 (#2692): it never follows a redirect now, and reads a 408, 409, 410, 422 or 507 with no code of its own as
      timeout, invalid, not_found, invalid and limit, where it answered vendor for all five.
    • Live view (#2768): screen.watch on a Linux desktop needs the X server's DAMAGE extension and is refused
      without it. E2B, Agent37 and Orgo desktops, and Xvfb, have it.
    • Timeline readers (#2747, #2762): a computer_calls record always carries started_at; computer_act and
      computer_observe rows no longer show typed text, a click's point or a scroll's direction; callWords and
      stepWords answer the past tense. A reader that built its own start time or matched the old words reads the new.
    • A run's own computer can be code-only or tools-only, and can have no shell (computer.surface, computer.exec,
      #2740); nothing changes for an agent that sets neither.
  • For developers:
    • The kernel's MCP door is removed (#2722): SessionSpec.mcp, SessionMcp, McpClient, AdapterCapabilities,
      AgentAdapter.capabilities(), AdapterFacts.mcp, ModelProxyOptions.mcp, mcpRoute, McpHandler,
      McpRouteOptions and McpCallResult.parked are gone; hand a session its MCP tools as host tools in
      SessionSpec.tools.
    • @parcha/agentrun-harness-claude no longer exports CLAUDE_WEB_SEARCH_USD; read NATIVE_SEARCH_USD from
      @parcha/agentrun-kernel/host (#2688).
    • The computer package: E2BSandbox.commands is now rpc, sendStdin, closeStdin and kill (#2716); the E2B and
      Modal SDK slice types are replaced by the SDKs' own (#2694); HostFacts.labelRules takes maxBytes and measure
      and tagsOf and labelsOf are gone from dist/modal.js (#2712); Daytona's listLabelled, probe,
      daytonaCommands, DaytonaCreateOptions, its timeouts and dist/conformance/daytona-sweep.js are gone (#2624);
      Agent37Error no longer carries status (#2692).
    • A computer MachineHost has twelve members (#2793): prepare and testConnection are gone (a keys test is one
      list, which must create nothing), setTimer and setLabels are one optional extend, id is descriptor, a
      provider's mounts and capabilities are read from descriptorOf(id) (the MOUNT and CAPABILITIES exports are
      gone), hostedProvider(...) is new HostedProvider(...), and Agent37Error and Agent37TunnelError are gone (a
      tunnel failure is an SshTunnelError). The changelog lists each with what to do.
    • On the shared machine-host adapter, every MachineHost member answers its value or throws (a HostError(code, message)), never a Result; exec takes no { stdin } flag; Machine is { id, labels, state, endsAt? } with
      state one of starting, running, stopped, gone; prepare, stop and the stop facts are gone, and a hosted
      computer is deleted, never stopped and kept (#2792).
    • GET /agents/{agent_id} and its revisions serve the pin as mcp_pin: per server, the pinned tools and their prices,
      never the server's URL (#2719).
    • computer-bench (a private package) renames fields in report.json, counts a computer call as a step, and requires
      each needs.apps entry (#2778, #2739, #2741).
    • Template API: each template in GET /v2/templates has id, bundled, source, sha,
      publisher, category and tags; an agent applied from one carries template: {id, version, sha} (#2728, #2732).
      GET /v2/runs/{run_id} serves input and file_ids (#2774), and result.output_error is new on the TypeScript and
      Python clients (#2725).

What you can do now

  • Install a template from GitHub: agentrun template install github:<owner>/<repo>/<path>@<ref> [--dry-run] [--trust],
    or Templates → Add from URL, and read the package before it is stored (#2728, #2789).
  • Give an agent a typed input and get a run form drawn from it; a bad input is a 422 with one issue per field (#2731).
  • Set a list in a template (type: array) and have it reach each scheduled run as a typed list (#2733), and filter a
    template's webhook trigger by a parameter, so a Composio project's one URL starts a run only for your repository
    (#2757).
  • Ask a person a question from a Claude Code or Codex agent (#2743).
  • Find a tool, call it by name or script several calls from any harness (#2749), and see Settings → Tools list each
    saved MCP server's tools with the permissions the server offers (#2764).
  • Open an app on a Linux desktop, and let an agent do it (#2727, #2735).
  • Give a desktop or a phone a code-only or tools-only surface, or no shell (#2740).
  • See what the cost of a run reads like on every harness, and which tool calls were not priced (#2761).
  • Let web_fetch read a site's atom, rss and JSON-LD feeds, such as a repository's releases.atom (#2773).
  • Read why a saved MCP server's tool can't be granted, on every harness it fails on (#2738).
  • Grant a monid endpoint by name, with its price held on the run's ledger before each call and named on the charge
    row ("Tool cost: monid charged $0.0260 for monid:apollo/people/match"); a re-priced endpoint is refused unsent
    (#2765).
  • Fence an MCP tool's object argument to named keys in a pin, so an unattended CRM write sets one field and nothing else
    ({"keys": ["hs_lead_score"]}), also from a template parameter (#2798).
  • Read why a learned workflow is held when its replay could not run an MCP tool: the hold names the tools with no
    readOnlyHint, and the Learning tab says what a promoted workflow was decided on (#2801).
  • Replay a learned workflow on graded runs before it serves, and read each decision at
    GET /agents/{id}/learning/promotions (#2644); grade a run with PUT /runs/{id}/grade.
  • Run an agent's computer use against a benchmark with computer-bench, now with a Linux desktop suite (#2577, #2744).
  • Turn on Allow reads for a saved MCP server: its read-only tools run with no one asked, its writes keep asking, and
    a destructive call takes a second click to allow. A tool's effect comes from the server's annotations, and a tool with
    none still asks (#2782).
  • Grant monid endpoints in the console's agent editor (Tools step, monid:<provider>/<endpoint> or
    monid:<provider>/*), with the server's own words for what a save left out or refused (#2822).
  • Let a learned workflow's replay record an escalate that reads a judge's field instead of stopping on it, and let its
    refit switch an escalate or a guard off; a replay that hands back every input is held (#2815).

Fixed

  • Runs:
    • A run starts sooner: a runner in the server's process claims a new run as soon as it is queued, and a first claim no
      longer asks the vendor to list older computers of the run (#2754).
    • A trigger's message that reaches a run just as it comes to wait on a person is delivered after the person's answer,
      where before it was refused and its firing still read sent (#2769).
    • A code-mode script replayed after a restart records each computer call once (#2755).
    • A CLI look-back cut by a lost runner is asked again once, not twice, and its spend stays the look-back's (#2668).
  • Spend:
    • A capped Codex, OpenCode or pi run's input estimate counts text at its own bytes, and on a route straight to OpenAI
      counts the input exactly (#2772).
    • A capped run on Codex or Claude Code prices its next request from the cache its previous request left warm, so a
      Codex run under $0.50 is no longer told to finish with $0.45 left (#2786).
    • A Codex or Claude Code run under a small cap (say $0.50) now finishes a job of a few cents, including after an
      approval or a deny (#2791, with #2786). A run resumed within 4 minutes, or a request sent after Codex's own side
      request, is priced from the cache. A request leaves room in the cap for the tool calls it starts while it streams.
      The cap still holds: a run never spends past it.
    • Behind serve's gateway, pi sends a Claude model on Anthropic's Messages API, so its prompt cache works: the second
      turn of a run reads the prompt from cache where every turn paid full input. The gateway must serve /v1/messages
      beside /v1/responses (LiteLLM's proxy and OpenRouter do); other models stay on the Responses API (#2799).
    • A CLI turn cut by a park (a host tool call held for your approval) is booked at the usage its provider reported,
      where the run's cost read unpriced (#2737).
    • A monid call's approval card shows its price: the hold at the endpoint's pinned price (#2687, #2765).
    • A paid call books on the run's one ledger or is not sent: a monid call, Parallel search and extract, and the browser
      refuse a call with no run to book under, and a learned workflow's web nodes book on the run's ledger (#2714).
    • The per-process spend engine is deleted; the run's ledger is the only count (#2724).
    • A run's usage row reads the same on every harness: "Cost: …" once a run ends, "Cost so far: …" only while it goes,
      Claude Code, Codex and OpenCode count tool calls with no price, and a failed request with no price says its cost is
      unknown (#2761).
    • An agent's cost per run counts only its own runs, and a run's cost and path land under load (waits up to 30 s, not 2
      s); the series says how many runs' paths are not known (#2642).
    • A long answer through an OpenAI Responses gateway that writes its JSON with spaces (LiteLLM) is priced; a request
      that answered while its runner's claim was lost is booked at its cost when the run resumes, also in a learned
      workflow; a run cancelled while its first request is being admitted never sends it and costs $0; and an unknown
      cost counts at what the cap held for it, never as a full cap, for the grade and the look-back (#2814).
    • A monid call that monid charges more than the run held for it stops the run at once, booked at what monid charged
      (#2818).
  • Tools:
    • A call waiting for your approval whose tool you then set to deny is refused, and the run says why; a call you allowed
      shows what it did and its result, with "(ran once you allowed it)"; on Claude Code and OpenCode tools_call and
      codemode each have their own row (#2764).
    • A script on pi reads a nested call's JSON result as a value (#2767).
    • web_fetch reads feeds (application/atom+xml, rss+xml, ld+json) as text (#2773).
    • Claude Code follows the tools_search / tools_call switch: with either set to deny, its own tool search is off and
      every tool is listed. Settings → Tools lists tools_search, tools_call and codemode, and codemode's description
      names real functions of the run and says to leave off a harness's prefix (hands_, mcp__hands__), which failed
      OpenCode scripts (#2795).
    • Deleting a Composio preset server in Settings → MCP servers also deletes the server the preset made in your
      Composio account; if Composio refuses, the answer and the console say what to remove there yourself. Saving a
      Composio server whose picked actions Composio serves none of answers 422 naming them, not a bare 424 (#2788).
    • A saved MCP server's credential is sent only to the address it was saved with, even when the owner moves the server
      to a new address with a new credential while a listing waits (#2719).
    • A server started by agentrun-serve (and so npx agentrun) can start a worker thread, so a codemode script, and
      pi's code mode on a desktop or a phone, no longer fail on --input-type (#2749).
    • A call that waited for approval reads in order on every harness: the call, "Asked your approval" at the time it was
      asked, then the result (#2820).
    • agentrun-serve eval-template runs its agents with approvals never, so a template whose tools ask is measured
      instead of stopping at its first ask, and the report says so (#2816).
  • Templates: a required text field sent empty is refused as missing, so the run form says "Company is required"; a
    source with a % that starts no escape is refused by a sentence that names it; an install refusal shows once (#2789).
    market-monitor's digest header starts at the window's first date on every harness (#2811).
  • Computers:
    • The Screen tab shows who drives and the first frame when they happen (400 ms reads while a change is on its way,
      not up to 3 s late) (#2730), reads right at a computer's start and end, has one Take over, and a phone's bar wraps
      at a phone's width (#2745, #2736).
    • A Mac's System Settings steps settle in 1.1 to 2 s, where each waited out 3 s (#2734), and Computerd settles on
      accessibility notifications (#2756).
    • The live view sends frames on screen change, up to 15 a second, and counts its cap from the start of each frame's
      read (#2758, #2763).
    • Agent37: every exec on an instance shares one SSH connection, so a step no longer opens a tunnel and a handshake of
      its own (#2751, #2695).
    • The exec pre-check names the deadline a write ran out of, and a set waits its full hold (#2726, #2577).
    • Settings → Computers says a Mode 2 vendor sandbox (Modal, E2B, Agent37) isn't ready, in plain words, instead of
      showing it ready and failing at its start; an agent's save warns, and a run's start refuses it with 409
      agent_not_runnable before any computer is made (#2796, #2804). In Mode 2 an E2B or Agent37 desktop is not ready
      and refused the same way (#2812); Orgo's desktop, a phone and a browser run as before.
    • The live view keeps moving while an input batch waits; a script on a vendor desktop may start a process
      (computer.processes.start) under the computer's own defaults; a failed computer row reads "Could not take a
      screenshot: …", and a no-ui computer_observe row names each call it made (#2806).
    • A computer deleted mid-run answers not_found, also for a call in flight (#2800), and a desktop or phone deleted
      under a live run reads "This desktop has been released." on its Screen tab, with Take over gone (#2817).
    • A desktop with no ui group (E2B's, Agent37's) can be driven by the tools alone: computer_observe and computer_act
      with app open that app first, its background is a slate colour, not black, and an ended run's computer reads
      session: "closed" (#2796). Agent37's desktop recipe needs agentrun-box rebuilt once.
  • Learning: the learned workflow's replay, guards and money rule are in Highlights (#2603, #2639, #2644, #2641); the
    learning batch restart test no longer ends now and then with the batch stopped: cost_unknown, because its runner
    drains at the stop as the CLI's does (#2781). A learned workflow held because its replay could not run an MCP tool says
    which tools and why, and a dry-run promote is no longer shown green (#2801). The Learning tab says "replayed 20", not
    "replayed 20/0", for a promotion that names no replayed runs (#2808).
  • Learning, a grade's words (#2648): PUT /runs/{id}/grade takes label (the grade's words) and answers feedback.
    A wrong grade of a run its learned workflow answered opens the same run again on a revision ask for that workflow;
    a run the agent finished, graded with a label, opens again on a look-back. Neither round answers the job again,
    and neither counts as a new run. A run's path is who wrote its final answer, and the cost per run counts every round.
    After such a round the live Timeline shows the round's rows, then the learning row, as after a restart (#2824), and
    the Learning tab's chip text comes from the server's words alone (#2830).
  • Server: on Postgres, a store's pool connections are named agentrun-store:<schema> in pg_stat_activity, as its
    lease and owner connections already were, so an operator can tell which store holds each one (#2770).

Also in this release

  • Computers:
    • Computers' shared adapter takes more from each vendor: Daytona runs on it (#2624), and its list, stop and check
      contracts (#2691), label byte budget (#2712), REST client and page loop (#2692), stream, bound, not-found and value
      helpers (#2689), SDK types (#2694), Agent37's SSH exec (#2695) and the window manager as a host fact (#2784) are in
      it; after an outside review, a provider's descriptor is the one source of its facts and each step has one path
      (#2793).
    • Exec is a byte-exact two-way stream with a pre-check a new vendor runs first (#2716), the daemon version matches
      exactly and builds without Go (#2715), and isolation fails on any listener in the user's network that is not
      computerd's (#2723).
    • A computer run's call list and Mac movies: a script's calls and what each acted on (#2747), draw marks on a Mac's
      live view and cut a movie per demo (#2721).
    • PROVIDERS.md is rewritten for a vendor's engineer, says how to contribute a provider and when you need Go (#2693,
      #2686, #2720), and follows the host contract above (#2827).
    • The release workflow publishes the box image a remote computer with an Archil mount starts from,
      ghcr.io/parcha-ai/agentrun-box:<recipe hash> and :<release tag>, and builds a public E2B template from it once
      the release environment holds an E2B key (#2807). The image carries its desktop, one recipe for every provider
      (#2828): this release's is ghcr.io/parcha-ai/agentrun-box:7f2fb1acc3fa.
    • Computerd: one copy of each input gesture loop (#2685), dead flags and unused codes deleted (#2683), the conformance
      cases run directly and the ui cases first (#2684, #2697), and the Mac mini runbook moved out of the package (#2682).
  • Learning: the compile turn gives a step a guard that starts at threshold 0.7 and, when the run can call Jev,
    both or jev (#2641).
  • Tools: one list for the table, askable, replay and blurb names (#2625); one native web search fee constant,
    $0.01 (#2688); max_tool_calls counted every nested call (#2760), until #2771 removed the budget.
  • Docs: a Composio quickstart in the README (#2717); an AgentRun ship lifecycle and native stack workflow (#2750).
  • Clients: generated client files no longer carry a hash header (#2797), and an operation that documents 204 beside a
    2xx body is typed as the body or nothing (X | undefined, X | None) (#2788).
  • Dependencies: none.
  • Tests and cleanup: the Composio fake counts only the requests on its own path (#2776); a PATCH that moves a
    schedule's time fires it at the new time, tested on the fake clock (#2779); every Postgres test closes its stores and a
    connection left open fails the test by name (#2770, with the settled flaky tests in #2769 and #2781); the store's
    transaction deadline test checks its own two timers, and a timing-only probe test is deleted (#2813); a test fake
    keeps only its own requests (#2819); the workflow-first restart test no longer asserts an exact claim count (#2826).

Known issues

  • Mode 2 (an Archil workspace) can't start E2B, Modal or Agent37 sandboxes, or E2B or Agent37 desktops, yet: they need Archil's image in your vendor account. Settings shows them as not ready, and a run that asks for one is refused before anything starts. Mode 1 E2B desktops and Docker work.
  • Agent37 desktops need the agentrun-box template, which exists only in AgentRun's own Agent37 workspace today, so a desktop in your own Agent37 workspace may not start in either mode. A fix is coming.
  • On Codex, a run whose cap is under about $2.50 can pause after spending about $0.20 (the cap holds a request's largest possible output). A fix is coming.

Known limits

  • The learning loop is further along but not finished (see Highlights); a learned workflow that answers wrong with
    confidence is not caught by the money rule.
  • A Codex run has no "Ran a script" row for its code mode yet; its calls show as their own tool rows.
  • Daytona takes no Archil mount until Archil's client works there.
  • deploy/compose.yaml (a server and runners) has no Docker socket, so its agents get file tools and web_fetch but no
    shell; for a Docker sandbox, use the root compose.yaml, run the server on the host, or use a Daytona or E2B computer.
  • The other limits listed for beta.32 still hold.