Skip to content

AgentRun v0.1.0-beta.42

Pre-release
Pre-release

Choose a tag to compare

@claudio-michel claudio-michel released this 17 Nov 14:18
· 246 commits to main since this release
3b10f6f

AgentRun 0.1.0-beta.42

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 learning you can check and computers you can watch. A learned workflow is
compiled once and promoted only when it agrees with the agent's own answers, and it compiles to a cascade that answers
from compact facts. An agent's desktop or phone can record a video of each run's screen, which the API serves and the
console exports, and a run's phone can carry real apps. Agents read the owner's calendar through the new
calendar_events tool, and a feed trigger on a JSON list keeps only the fields it names and fires again when an item's
date moves on. Computers gain boxd, Morph and Fly Machines providers, and E2B and Modal machines can dial the server.
The store does not change: SQLite keeps schema version 11 and Postgres version 16, so going back to beta.41 is changing
the image back on both. The launcher is @parcha/agentrun 0.4.10, the Python client parcha-agentrun is 0.3.10, and
the server image is ghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.42 (linux/amd64 and linux/arm64).

No fresh store: beta.42 opens the data beta.41 left. Keep the same data folder and the same start command, and
change only the image to ghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.42. Your agents, triggers, runs and settings
are all still there.

  • No store change. SQLite stays at schema version 11 and Postgres at 16 (0016_minted_tokens), with no migration
    and no new table or column. At each start the server keeps its own addresses in one settings row
    (mcp.server_own), which beta.41 ignores.
  • Going back is setting the image back to v0.1.0-beta.41, on SQLite and on Postgres alike: beta.41 opens the
    store beta.42 used, with every agent, trigger, run and setting in it, and no pg_dump restore is needed. What
    beta.41 does with what only beta.42 writes:
    • An agent whose learning is an object: beta.41 reads it back as saved and runs the agent, with learning off. It
      refuses to save a new definition of that agent (422 wrong_type on definition.learning, "true, false or
      "notes""): to edit it on beta.41, set learning: true, and set the object again once you are back on beta.42.
    • The calendar: beta.41 has no calendar routes (404 "no route /settings/calendar") and no calendar_events tool.
      The saved address stays sealed in the store, and beta.42 reads it again on return: there is nothing to save again.
    • A feed trigger on a JSON list with fields or updated: beta.41 keeps it on and shows both pointers, but uses
      neither. A new item fires with the whole item cut at 16 KB (item_truncated), and an item whose date moves on
      fires nothing. Back on beta.42, each item first seen on beta.41 starts one more run at the first poll, as
      update: true: a target that must act once per item should skip an update it has already handled.
    • Run recordings: beta.41's run read has no recordings, and GET /v2/runs/{run_id}/files/recordings/<computer>.mp4
      answers 404 "no such file". The files stay beside each run's folder, and beta.42 lists and serves them again on
      return.
  • Learned workflows are promoted by agreement with the agent's own answers, not by grades. To compile sooner or
    hold drafts to a stricter bar, set learning: { compile_after: 5, agree_min: 0.9 } on the agent. A client that reads
    definition.learning as a boolean or "notes" must also accept an object, and treat it as on.
  • The compile turn asks for no step guard and refuses a draft with one. Learned workflows already saved keep
    running as they are; a revision of one drops its guards.
  • A workflow agent's runs no longer look back on their own: give feedback (a grade with words, or a follow-up
    message) on the runs it should learn from. A client that read feedback: "revise" on a grade reads null for a
    served run now.

Highlights

  • Learned workflows are promoted once, by agreement with the agent's own answers. An agent's learning takes an
    object, { compile_after?, agree_min?, alone_min? }, which is on with those numbers set; each one left out is its
    default (20, 0.8, 0.7), and true is every default. The first compile is asked once the agent has compile_after
    completed runs of its own, no longer on a count of kept notes. A new revision's first claim replays it on the inputs
    of those runs and scores each against the agent's own recorded answer, not a grade: it is promoted (source: "check") when it answered at least alone_min of them alone and at least agree_min of those said what the agent
    said. Under either bar it is held, and the agent's next run asks one revision with the inputs it disagreed on; held
    again, it stops (check.stopped), the agent keeps its runs and nothing compiles again. A draft with nothing to check
    it on stops with a reason that starts "nothing to check it on". The promotion record carries check (the bar, n,
    agreed, alone, the runs read, skipped_tools, and on a hold the disagreements), later grades and refits don't
    move it, and a newer revision waits for a running check (#3032).
  • Compiled workflows take the cascade shape. The compile turn's worked example is the cascade a repeated job
    compiles to: a call step fetches the page, a fast-tier extract step copies compact facts from it, the answer step runs
    on the agent's model from those facts and its own draft alone and re-picks once, in a loop of 2, until its quotes pass
    quote_missing, and an escalate step hands back a second miss or a case the policy does not cover. A draft whose
    model step off the fast tier reads a fetched page (through its state map, or the whole state with no map) is
    refused in one line naming the step, and the quote check's re-pick asks for the compact state map where it read
    every page on the agent's model. In the DSL, a whole-string placeholder marked optional, "{draft?}", is null while
    its path is absent; an unmarked one still fails missing_interpolation. placeholderValue(state, path) and
    placeholdersOf(text) are exported for a host that checks what a step reads (#3025).
  • Run video. definition.computer.record: true (off by default) records an agent's desktop or phone from the
    claim's attach until the computer's release or the run's end, with or without a viewer, as
    recordings/<computer>.mp4 beside the run's work/: H.264 in fragmented MP4, at most 10 frames a second and 1280
    pixels on the long side, no audio. A park pauses it and the resumed run goes on in a new file. It never holds up the
    live view or the agent, and a cap (500 MB across a run's files, 2 hours), an encoder that stops or a screen that stops
    answering ends the recording with a line on the timeline, never the run. The video shows everything on the screen, a
    person's typing during a take-over and any secret included: pixels can't be redacted. A server without ffmpeg on its
    PATH refuses record: true at save, as does a computer that is not a desktop or phone (422 not_allowed) (#3047).
    The run read lists recordings (path, computer, seconds, bytes, finished), GET /v2/runs/{run_id}/files/recordings/<computer>.mp4 serves one as video/mp4 to whoever may read the run's files
    (?download=1 for an attachment), and the files route answers one byte range for every file, so a <video> can seek
    (features.files.range: true). A recording reaches the disk as it is made, a fragment every 2 seconds or so. The
    console's Screen tab says "Recording" while one is written and offers Export once it is finished (#3063).
  • A calendar for agents. An owner saves the server's one calendar by its secret iCal address (Google Calendar's
    "Secret address in iCal format", Outlook's ICS link, iCloud's public link), and the new host tool calendar_events {from?, to?} reads its events. The routes are GET, PUT and DELETE /v2/settings/calendar and POST /v2/settings/calendar/test, for the owner's signed-in session alone (the token is 403); the console's page comes
    next. PUT {url} takes https and reads webcal:// as https; plain http is taken only to a loopback name
    (localhost, 127.x.x.x, [::1]), which only a local install reads. The address is sealed in the store and is in no
    answer, event, log line, error or run input, with its secret token and any address it redirects to scrubbed too. The
    test answers in words ("Work: 214 events, time zone America/New_York, next: Standup, Mon Nov 17 09:30") and saves
    nothing. The tool answers {calendar, time_zone, from, to, count, truncated, events}, at most 200 events sorted by
    start, over now through 7 days unless told otherwise and 62 days at most. Repeating events are expanded in the
    window (RRULE, RDATE, EXDATE, moved and cancelled instances, RANGE=THISANDFUTURE), each event keeps its own time
    zone across a DST change, and a read's work budget (15,000 rule steps, 1 s) answers a calendar that needs more in
    words, never as a partial list. It runs on every harness, is allowed by default and not metered, and an event's words
    are untrusted, as a page's are. The fetch is held to the server's one address rule, an MCP server's: a private
    address only when its host is allowed in Settings → MCP servers, link-local and this machine never, loopback only on a
    local install and never at the server's own ports, every redirect checked again; 10 MB, 20 s, 5 redirects, a read
    reused for 5 minutes and then revalidated. A template's tools may name calendar_events, and a template eval case
    may say calendar: <file>.ics (#3012).
  • Real apps on a run's phone. The operator names APK files on the machine that drives the phones in
    PHONE_HARNESS_APKS, a comma list of <absolute path>@sha256:<hex>, and each phone installs them before the agent
    starts on it. A file that is not there, or whose bytes are not its pin, refuses the claim in words before a phone is
    rented. The timeline names a phone's Clock, Contacts and Files apps and the Wikipedia app by the names on their icons
    (it said "Deskclock" for the Clock) (#3037).
  • A JSON-list feed keeps what it names and sees updates. A feed trigger reading a JSON list takes fields, 1 to 50
    JSON Pointers into an item (["/number", "/title", "/user/login"]), which keeps only those in the payload's item
    before the 16 KB cap, so a GitHub pull request now fires without item_truncated; and updated, a pointer to the
    item's date ("/updated_at"), which fires it again with update: true when that date moves on. Every feed pointer
    reads an item's own keys only (never __proto__ or constructor through the prototype), and a feed's seen set tells
    a missing date from a real one, so an item dated 1970-01-01 or earlier fires its update; a cursor saved before this
    reads a JSON list's recorded 0 as a missing date, once. An eval's feed fake matches an item by the trigger's id
    pointer, and validate-template refuses an eval case's web: fixture with no search in eval-template's words
    (#3041).
  • More computer providers. boxd (providerFor("boxd", { apiKey })): each computer a boxd KVM microVM running
    Ubuntu 24.04, as a sandbox or a desktop, on the shared machine-host adapter, with network none as the machine's
    egress allowlist, the janitor's lifetime and boxd's own delete timer at the adapter's backstop; Settings takes
    BOXD_API_KEY, with BOXD_ORG and BOXD_REGION optional. A boxd computer takes no mount yet, so serve refuses a
    boxd sandbox and runs a boxd desktop in either mode (#3035). A boxd desktop starts from a snapshot,
    agentrun-desktop-<recipe hash>, that prepareImage() builds once in the key's org with Xfce's text editor, so
    apps.launch has an app to open; a desktop on boxd's own image, or with vcpu and memory, is refused in words
    (#3045). Morph (providerFor("morph", { apiKey }), @parcha/agentrun-computer/morph): a kind linux instance
    started from a snapshot, its lifetime Morph's TTL, every command over SSH through Morph's gateway with the instance's
    own key, so the account's key never goes over SSH; no Archil mount or desktop yet, and serve reads MORPH_API_KEY
    (#3038). Fly Machines (fly, @parcha/agentrun-computer/fly): a dial host whose machines boot the box image and dial
    the server, with FLY_API_TOKEN and FLY_APP in Settings → Computers, and packages/serve/scripts/dial-conformance.mjs --provider fly runs its Core inside the server the machines dial (#3042). E2B and Modal computers can dial the server
    with connect: "dial" (e2bConnect, modalConnect); exec stays the default, and on dial both take kind linux alone
    and no Archil mount (#3040).

Breaking

  • @parcha/agentrun-computer: the phone-harness provider's apks option takes { path, sha256 } entries, not bare
    paths (#3037).
  • @parcha/agentrun-computer: a provider can no longer declare labels: "machine". HostFacts.labels takes { count, bytes, packed } only, and Machine.name, Machine.image, HostedOptions.image and the sweep --image flag are gone;
    a vendor with no labels of its own waits for a first provider that needs the mechanism (#3034).
  • API: definition.learning may be an object, { compile_after?, agree_min?, alone_min? }, as well as true, false
    or "notes" (#3032).
  • API: a wrong grade of a run its learned workflow answered opens no feedback round: PUT /v2/runs/{run_id}/grade
    answers feedback: null, and the revise round is gone (#3028).
  • Learning: the compile turn refuses a draft with a step guard (guard, its threshold and signal) (#3025).
  • Runs: a folder an agent makes named recordings at the root of its run's work/ is no longer listed or served: that
    path is the run's recordings (#3063).
  • computer-bench: mobileworld.adjust-brightness-max and mobileworld.adjust-brightness-min are removed (#2746).

Behaviour changes

  • A learned workflow is asked for a revision only when its promotion record holds it, once per hold (#3028). A
    hand-back, a stop by the money rule and a wrong grade of a served run no longer each ask a compile turn: the hold
    reads each among the workflow's other cases. The money rule still stops a workflow that costs more than the agent;
    the stop asks for no compile and the agent takes the run. An agent with learning: true asks no look-back turn
    unless its owner gave feedback on the run (a grade's words, or a message after the run's first); a run without it
    spends $0 looking back and keeps its learning dropped empty. learning: "notes" runs look back as before, and an
    SDK agent with learning: true learns through its compile turn alone.
  • A phone's live view comes faster and in smaller frames (#3039). phoned reads the screen every 200 ms while a
    screen.watch call waits, where it read every second, and a read whose pixels are the last one's is neither scaled
    nor encoded. The phone-harness provider's watch() asks for JPEG frames 720 wide when its caller names no width,
    about 2.2 times fewer pixels on a 1080 × 2400 phone. The Screen tab and the run video take these frames; input
    coordinates stay the screen's own pixels, and a caller's maxWidth still wins.
  • A runner in its own process honours AGENTRUN_MCP_ALLOW_LOOPBACK=1 from its own environment (#3049), as the
    server and the runner inside it already did; before, it refused every call to an MCP server on loopback. Each process
    reads only its own variable, and loopback means the machine or container the runner runs on. Such a runner never
    reaches a server's own port: 7340, the port of a loopback AGENTRUN_PUBLIC_URL, and the addresses the server keeps at
    each start (mcp.server_own), read at each claim. It says on stderr at its start that it takes the variable. A run of
    a revision saved before pins whose MCP server is on loopback, on a server started without the variable, now gets a 409
    that names the cause and the variable, where it said to save the agent again.

Known issues

  • The Cloudflare provider in this release fails its live conformance run: its relay answers a create with a bare 500
    when the container is slow to start. Don't use it on beta.42. The relay fix is in beta.43 (#3072), not in this
    release.
  • The browser screencast is not recorded yet; run video covers desktop and phone.
  • Going back to beta.41 and returning: a JSON-list feed trigger with updated starts one repeat run, as update: true,
    for each item first seen on beta.41 (see the going-back notes above).
  • Right after a release is published, an install can fail for up to about 10 minutes, until every package is visible
    on npm: a package whose dependency npm does not show yet fails to install. Wait and install again.
  • A token's GET /v2/settings/calendar is refused 403, as it should be, but in the words for saving a secret; every
    calendar route takes the owner's signed-in session.

Dependencies

  • @parcha/agentrun-serve adds ical.js 2.2.1, pinned exact (MPL-2.0, no dependencies of its own), the calendar's
    parser; it loads at a calendar read only, never with the server (#3012).
  • The server image's ffmpeg is one static binary of about 6 MB built for the screen recorder alone (PNG and JPEG frames
    in, H.264 in fragmented MP4 out): FFmpeg 9.0.2, x264 and zlib, each pinned by sha256. The image is 163 MB smaller than
    with Debian's ffmpeg and 2.8 MB larger than with none, and its build fails unless the recorder makes H.264 files
    from both kinds of frame. It is GPL, as Debian's is; the licence files are in /usr/local/share/doc/ffmpeg (#3056).
  • The repository root adds @boxd-sh/sdk 0.3.1 as a dev dependency for the boxd provider; a host that runs boxd
    installs it, and @parcha/agentrun-computer depends on nothing new (#3035).

What you can do now

  • Let a learned workflow serve only once it agrees with your agent's own answers, and set when it compiles and how
    strict its check is (#3032, #3025).
  • Record a run's desktop or phone, seek through it in the browser and export it from the console (#3047, #3063).
  • Give your agents your calendar (#3012).
  • Install your own apps on every phone a run rents (#3037).
  • Watch a repository's pull requests whole, and start a run when one is updated (#3041).
  • Run computers on boxd, Morph or Fly Machines, or have E2B and Modal machines dial the server (#3035, #3045, #3038,
    #3042, #3040).
  • Point a runner in its own process at an MCP server on its own loopback (#3049).
  • Start the launcher with as many secrets as you have (#3052).

Fixed

  • The launcher refused to start once the secrets it hands over came to more than 4,096 bytes, which a dozen real
    keys or one RSA AGENT37_SSH_KEY could reach. A fork of its shell now writes the pipe as the server reads it, and no
    secret goes into an argument, the server's environment or a file, as before (#3052).
  • apps.launch with a url into a Linux desktop app that already runs answered "opened" when the launch's launcher
    ended at once and left a process that showed a dialog a moment later. Computerd takes the launch's guard from the
    start itself, and the call refuses, naming the dialog, and ends it (#3048).
  • On Android, a compact ui.observe took a list that is clickable as well as scrollable (Clock's cities) for one row
    named after its first text, and hid its rows; it lists them now (#3065).
  • Morph's box build timed out on its own base snapshot: POST /snapshot answers only once the snapshot is made (78 s
    measured), so it now has the build's remaining time. A build looks for a base by its label first and again after a
    lost create, so a lost create no longer leaves a base snapshot nothing deletes (#3053).
  • Morph's first SSH exec failed with "Cannot parse privateKey: Unsupported key format": the PKCS#8 key Morph answers
    is converted before the login (RSA to PKCS#1, EC to SEC1, Ed25519 to OpenSSH's own format), and an Agent37 sshKey
    in PKCS#8 the same way (#3054).

Also in this release

  • Computers: a Cloudflare provider (@parcha/agentrun-computer/cloudflare, id cloudflare), a Linux sandbox in a
    Cloudflare Container whose computerd dials the server, through a relay Worker the operator deploys once from
    packages/computer/relays/cloudflare (Workers Paid) and whose URL and token go into Settings as
    CLOUDFLARE_RELAY_URL and CLOUDFLARE_RELAY_TOKEN; see Known issues before using it (#3046). Computerd.app can be
    notarized with scripts/notarize.sh, and scripts/bundle.sh signs from a keychain of its own
    (COMPUTERD_SIGN_KEYCHAIN) (#3044). The provider scorecard adds boxes.dev as two rows and re-words Cloudflare's row
    for the dial path (#3033).
  • computer-bench: android-dev gets 17 new tasks, eight dev and nine heldout, each verified on a cloud phone, with
    UI-only oracles; every exec-deny phone oracle first waits for its app's window, and the suite README shows how wide
    the paired-difference interval is at the suite's 37 tasks (#2746). The Android tasks done in Settings, Clock, Contacts
    or Files (16 in android-dev, 6 in mobileworld) deny the agent exec and do the task through the UI in their oracles,
    and device-name, battery-level and alarm-time have an oracle now (#2748).
  • Docs: the setup skill (docs/SKILL.md) checks that the person can reach the machine before a local install: in a
    hosted session whose loopback their browser cannot open it stops and offers a session on their own computer or
    AgentRun Cloud, and after the server starts it waits until they confirm the console opens (#3057).
  • Tests and CI: the CLI tests give server openapi room for the whole OpenAPI document, which had grown past
    spawnSync's 1 MiB buffer, and a spawn that fails fails the test with its cause (#3059). A release takes about 20
    minutes and one approval instead of about 85 minutes and three: every npm layer publishes back to back with one wait
    for the registry, the image builds each architecture on its own native runner, and a gate job checks the tag first;
    PyPI, the box image and the E2B template can now be published for a tag whose npm publish then fails (#3036).

Known limits

  • The limits listed for beta.41 still hold.