Skip to content

AgentRun v0.1.0-beta.43

Pre-release
Pre-release

Choose a tag to compare

@claudio-michel claudio-michel released this 17 Nov 14:42
· 234 commits to main since this release
e6e0ed4

AgentRun 0.1.0-beta.43

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 learned workflows you can keep trusting. The old promotion loop is gone:
a learned workflow has one path to serving, its one check against the agent's own answers, then promote, or hold and
one recompile. Once it serves, a sample of its runs is answered again by the agent itself, and too many disagreements
hold it. The release also fixes the Cloudflare relay's bare 500 on a slow create and the CPU cost of importing large
Codex histories. The store does not change: SQLite keeps schema version 11 and Postgres version 16, so going back to
beta.42 is changing the image back on both. The launcher is @parcha/agentrun 0.4.11, the Python client
parcha-agentrun is 0.3.11, and the server image is ghcr.io/parcha-ai/agentrun-server:v0.1.0-beta.43 (linux/amd64
and linux/arm64).

No fresh store: beta.43 opens the data beta.42 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.43. Your agents, triggers, runs, settings
and learning folders 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, upgrading or going back.
  • Learning records from beta.42. A promotion record beta.42 wrote as queued or skipped is left out of every
    read, and its revision is checked again at the agent's next run. Fields this release dropped (source, replay,
    refit and the rest) are ignored. Nothing is deleted from the store: beta.42 shows those records again if you go
    back.
  • Delete guard from every learned workflow. beta.43 does not run a learned workflow whose workflow.json still
    has a step guard: each run says "the workflow DSL refuses it: NODE_KEY_UNKNOWN ... unknown key(s) for a extract
    node: guard", hands the job back, and the agent answers. The promotions read still says that revision is
    serving. An agent with learning history from before this release can also turn learning off and on (or delete its
    learning folder) to start clean.
  • Going back is setting the image back to v0.1.0-beta.42, on SQLite and on Postgres alike: beta.42 opens the
    store beta.43 used, with every agent, trigger, run and setting in it, and no pg_dump restore is needed. What
    beta.42 does with what only beta.43 writes:
    • An agent whose learning sets any of the audit numbers: beta.42 reads it back as saved and runs the agent. It
      refuses to save a new definition of that agent (422 wrong_type on definition.learning,
      "definition.learning.audit_every is not one of the loop's numbers: compile_after, agree_min, alone_min"): to edit
      it on beta.42, take the six audit numbers out, and set them again once you are back on beta.43.
    • Promotion records in this release's shape (no source, decided: "replaying", check.unstarted, a hold with
      audit): beta.42 reads them, and a workflow beta.43 held stays off its runs. Its serving entries have none of
      the five new fields, and its loop read answers with its own guards list, empty for a workflow with no guard.
    • Audits: beta.42 starts no shadow and has no audits route (404 "no route /agents/{agent_id}/learning/audits").
      audits.jsonl stays in the learning folder, and beta.43 reads every row again on return.
    • Shadow runs: beta.42 lists a shadow that beta.43 started as an ordinary run, in GET /v2/runs and in the agent's
      run list, with no audit_of, and the served run's cost.usd leaves the shadow's cost out. Back on beta.43 the
      shadow is out of the lists again and reads with audit_of.
  • Clients: stop reading guards, a promotion record's source, replay and refit, and the queued and
    skipped decisions; read check. A client that builds the promotions read's serving objects must add state,
    check, audit, audits_off and not_auditable, and a served run's cost.usd can be null while its shadow runs.

Highlights

  • A learned workflow has one path to serving. A new revision serves only once its one check against the agent's
    own answers promotes it; a hold asks one recompile, and a held recompile stops. Nothing else moves it: the threshold
    refit, the graded hold, the graded replay, the queue of revisions, rival settles, step guards, the Jev generality
    pre-check and the quiet window after a compile are deleted. A revision waits for a running check and records
    nothing. An owner's replay (POST .../learning/workflows/{name}/revisions/{digest}/replays) re-checks a revision on
    the agent's own completed runs: inputs: "held", the runs its last check ran on, or a number from 1 to 200 of the
    newest; with no earlier check it picks the agent's last compile_after completed runs. The console's cost card
    reads "Passed the agreement check: N of M agreed (bar 80%)" (#3064).
  • Serving workflows are audited against the agent itself. A run a promoted learned workflow serves first is drawn
    at 1 in learning.audit_every (10), then 1 in learning.audit_every_later (20) once its revision has served
    learning.audit_later_after (50) runs, seeded from the run's id. A drawn run starts a shadow run at its end: the same
    agent on the same input with learning off. Each finished audit is a row of audits.jsonl in the agent's learning
    folder, read with GET /v2/agents/{agent_id}/learning/audits (run, shadow_run, workflow, digest,
    promoted_at, source, agreed, served_cost, shadow_cost, turn, at), and an owner's grade of a run the
    workflow answered alone is a row too (source: "grade"). learning.hold_disagree (3) disagreements in the newest
    learning.hold_window (10) rows of the revision's current promotion hold it, and the agent's next run asks one
    recompile with those disagreements as data; a second hold within learning.stop_window (100) served runs stops it
    ("stopped: the workflow disagreed with the agent twice"). The shadow is no row of the run list; GET /v2/runs/{run_id} reads it with audit_of, and its cost is the served run's learning cost. A capped run's shadow is
    capped at what the run's cap left, so the two never spend past it, and a shadow never waits on a person: one that
    would ends there, parked, never counted toward a hold. Only runs started over the API are audited, and only for an
    agent every tool of which only reads, so a shadow never writes a second time; the promotions read says why one is
    not (audits_off, e.g. "audits off: send_email can write") (#3061).

Breaking

  • DSL: step guards are removed: a node takes no guard field, and GUARD_SIGNALS, GUARD_FIELDS,
    GUARDED_NODE_KINDS, SELF_CONFIDENCE_KINDS, the guardStep dependency, the guard escalation kind, the
    $confidence channel and softEscalates are gone; a replay obeys every escalate (#3064).
  • Learning: a promotion record keeps workflow, digest, at, decided, reason, check, runs, inputs,
    replay_id and commit; source, replay, refit, checked, since, held_on and skipped_by are gone, and
    so are the queued and skipped decisions (#3064).
  • API: GET /v2/agents/{agent_id}/learning/loop returns { agent, completed_runs, compiles }: guards is gone
    (#3064).
  • API: POST .../learning/workflows/{name}/revisions/{digest}/replays takes inputs: "held" or a number from 1 to
    200, and replays the agent's own completed runs, not its graded inputs (#3064).
  • API: the learning cost series' replay is the serving check's numbers, with agreed, alone and agree_min
    (#3064).
  • Configuration: AGENTRUN_SERVE_LEARNING_GENERAL_THRESHOLD is removed with the generality pre-check on Jev: the
    grader reads every note (#3064).
  • API: the promotions read's serving entries gain state (serving, checking, held or stopped), check,
    audit, audits_off and not_auditable (#3061).
  • API: definition.learning takes six more numbers, audit_every, audit_every_later, audit_later_after,
    hold_window, hold_disagree and stop_window, each a whole number from 1 to 1000 with hold_disagree at most
    hold_window, else a 422 on definition.learning (#3061).
  • API: while a served run's shadow is queued or running, its cost.usd, cost.all_in_usd and learning cost read null
    (#3061).

Behaviour changes

  • The compile turn asks only for a new workflow or one revision of a held one (#3064). A hand-back, a wrong grade
    and the money rule's stop ask for no revision. A compile whose drafts all fail their checks is a compile too: it
    writes a hold whose reason names the failed check, a failed first compile earns the one recompile, and a failed
    recompile stops; nothing is asked again for that revision. No new compile is asked while a draft that passed its
    checks waits on its grade.
  • A check's numbers add up (#3061). Its outcomes add up to its n: a replay run never created by the check's
    deadline is failed, and the record's new check.unstarted counts those. A check none of whose runs was created
    reads failed on GET .../replays/{replay_id}. A check whose every input stopped at a write stops with the cause
    and spends no recompile.
  • A promotion hold says whose it is (#3061): a hold with audit is an audit's, one without it its check's. The
    graded hold is gone: an owner's grade of a served run is one more audit row.

Known issues

  • The Cloudflare relay fix in this release (#3072) is not the whole fix: the relay needs #3090 too, so redeploy it from
    beta.44 (or main) once that ships. Don't depend on the Cloudflare provider on beta.43.
  • The browser screencast is not recorded yet; run video covers desktop and phone.
  • Fixed in beta.44 (#3099): a learned workflow the release's DSL refuses no longer reads serving; it is held, and
    nothing is needed on upgrade. On beta.43, a learned workflow that still has a step guard is not run while the
    promotions read still says serving (see the upgrading notes above): delete guard from it (#3064).
  • Going back to beta.42: it lists a shadow run beta.43 started as an ordinary run of the agent, and leaves the shadow's
    cost out of the served run's cost.usd (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.

What you can do now

  • Let a learned workflow serve only while it keeps agreeing with your agent, checked on a sample of its own runs, and
    read every audit (#3061).
  • See each learned workflow's state, its check's numbers and its audit line in one read (#3061).
  • Re-check a held or edited revision on your agent's own last runs (#3064).
  • Redeploy the Cloudflare relay and read every failure in words (#3072).
  • Import large Codex histories without the repeated work (#3076).

Fixed

  • The Cloudflare relay answered a create with a bare 500 when its container was slow to start. The create held its
    Durable Object over a call that could wait past blockConcurrencyWhile's 30 s limit, which reset the object. The
    idle timeout is now set outside that block, every route answers a failure as { error } in words, and the provider
    sends a create answered 500, 502 or 503 again (the relay answers a repeated name with the first's row). Redeploy the
    relay from this release (#3072).
  • Importing large Codex histories and discovering connector sessions did repeated work. Codex transcript
    reconstruction merges deferred user messages in one ordered pass, and connector discovery allocates session and
    project identities in one locked batch per pass, keeping durable canonical IDs and connection consent checks (#3076).
  • A run whose folder is on Archil and whose computer mounts it left its scratch behind. Its end removed
    .agentrun/tmp and .agentrun/output over Archil's file API while the mount held the folder's write delegation
    (409 OperationAborted), after waiting out a 60 s retry window. The run's end now removes the scratch through the
    computer when it mounts the folder; if the computer no longer answers, a Mode 1 run removes through its own folder as
    before, and a Mode 2 run's scratch stays with a log line saying why (#3058).
  • On Android, a compact ui.observe took a clickable list whose rows all fit on the screen for one row (Clock's
    cities) and hid its rows: phoned now knows a list by its class as well as by whether it can scroll now (#3073).

Also in this release

  • Computers: when computerd probe --sockets stops a root's walk at its entry budget, the reason names the three
    top-level folders with the most entries walked (/run: the walk stopped after 50000 entries, its budget (most under /run/x 41210, ...)), so a failing isolation.vendor-agent says what filled the budget (#3075).
  • Console: the Learning tab drops its step guard, refit, queued and skipped words, and the Replay button and cost
    card read the check (#3064).
  • Server image: the image build's ffmpeg check waits on the recorder instead of the clock, so the arm64 image
    builds under emulation too; nothing the image or the server does changes (#3067).
  • Release: server and workspace 0.1.0-beta.43, launcher 0.4.11 pinned to it, Python client 0.3.11 (#3084).

Known limits

  • The limits listed for beta.42 still hold.