Work collides.
TTL claims and FIFO contention replace polite assumptions with explicit, inspectable ownership.
Durable coordination / recorded office 01
Agents need an organization.
BATON turns long-running agent loops into a legible institution: scheduled routines, leased claims, scoped advisories, pinned objectives, recoverable checkpoints, and memory that must earn promotion.
03durable agent routines
42append-only source events
01real close / reopen recovery
0API keys required to reproduce
I / The organizational gap
Intelligence is not coordination.
Agent frameworks help a model act. BATON focuses on the institutional layer around those acts: who owns work, what changed, which warning applies, what survives a restart, and which lessons deserve to become policy.
Context drifts.
Versioned handoffs pin objectives, constraints, and delivered advisories at tool boundaries.
Memory lies.
Episodes can suggest lessons, but evaluation decides whether they are promoted, archived, or left undetermined.
II / The control surface
Read the office across time.
The primary interface borrows the grammar of an orchestral score. Each actor holds a lane; claims, advisories, checkpoints, and memory become distinct cues on one shared axis. Nothing is collapsed into a vague “agent activity” feed.
Recorded morning / logical time →
Replay complete
docs-agent
claim holder
deps-agent
recoverable run
sre-agent
advisory publisher
Schedule
A routine creates a versioned objective and budget envelope.
Claim
Work scopes have one owner token and a deterministic queue.
Conduct
Scoped advisories arrive only at explicit step boundaries.
Recover
Workspace and harness state resume from the last durable checkpoint.
Learn
Only evidence-backed lessons move into durable organizational memory.
III / Inspectable proof
The interface leads back to artifacts.
Claim queue, not choreography.
The fixture proves one holder, one queued contender, token-checked release, and automatic transfer on release.
Warnings become pinned data.
A scoped SRE advisory reaches the dependency agent at its next boundary and survives restart in the handoff.
The process actually closes.
The recorded run closes its SQLite
connection, reopens it, restores step 2, then continues. The
make reproduce-resume harness exercises 100 cycles.
Promotion is an evaluated event.
Candidate scores beat their registered baselines before the fixture lesson is marked promoted. See the complete event transcript.
IV / Installed control plane
The runtime is not trapped in the demo.
Install BATON to expose the same durable engine through a local HTTP service. Caller-owned loops register routines, coordinate work, persist tool boundaries, and resume after restart; the score can read those real events.
Register & fire.
Versioned routine specs become durable runs with pinned constraints, budgets, subscriptions, and declared work claims.
Coordinate live.
Claims, scoped advisories, objective changes, and usage signals use explicit API transitions with structured errors and an append-only event trail.
Restart & prove.
SQLite retains checkpoints and handoffs. Resume restores the last durable edge, while the evidence endpoint returns its exact versions and events.
V / Architecture
One compact, durable spine.
The MVP stays deliberately small: standard-library Python and SQLite, no model SDK, and no server needed for the published fixture. The optional local service exposes the actual runtime; the score remains its projection.
Routine spec
Objective, schedule, budgets, claims, subscriptions, and workspace ref.
Versioned handoff
Verbatim objective, constraints, and advisory set.
Tool checkpoint
Step, workspace reference, and harness state commit together.
Event ledger
Every organizational transition enters one ordered SQLite record.
Conductor score
Static Pages replays the fixture; the installed service can feed it current durable events.
VI / About & honest boundary
Small enough to inspect.
- It isA single-node runtime harness for deterministic, durable agent routines.
- It is not A distributed scheduler, agent framework, OS lock manager, or reliability benchmark.
- MeasuredJourney 0 behavior, event order, recovery state, and the registered memory-gate fixture.
- UnmeasuredFleet-scale throughput, cross-node consensus, and real-model quality.
The full score is interactive