DEV Community

Robin
Robin

Posted on

"Checkpoint Long-Running Agents Across Model and Effort Revisions"

#architecture #reliability #ai #testing

Checkpoint c17 was written under model revision M1 at medium effort. The service now defaults to M2 at high effort, while two tool schemas and the policy bundle also changed. Blindly resuming can replay an incompatible plan; restarting discards completed work. Compatibility must be a protocol decision.

Anthropic's July 24 Claude Opus 5 announcement says the available model improves long-running agent and coding work and offers effort settings. These are Anthropic claims, including its benchmarks and pricing comparison; verify current pricing and supported settings. This is the latest verified official signal, not a July 27 launch. Secondary July 27 statements without first-party verification are excluded.

This design does not re-explain a generic task state machine. It focuses on whether serialized reasoning-independent artifacts remain valid across model, effort, tool, and policy revisions.

Checkpoint envelope 

{
  "schema": 2,
  "operation": "op_91",
  "checkpoint": 17,
  "parent": 16,
  "model": "exact-model-id-at-write",
  "effort": "exact-supported-value-at-write",
  "policyHash": "sha256:policy-v4",
  "toolSchemaHash": "sha256:tools-v8",
  "baseRevision": "git:abc123",
  "completedEffects": ["effect_12"],
  "pendingIntent": {"kind":"apply_patch","inputHash":"sha256:..."},
  "evidence": ["test:unit:pass@abc123"]
}
Enter fullscreen mode Exit fullscreen mode

Do not serialize hidden chain-of-thought. Persist external facts, effects, revisions, and validation evidence.

Compatibility function 

def decide(cp, runtime):
    if cp["schema"] not in runtime.readable_schemas:
        return "REJECT_SCHEMA"
    if cp["baseRevision"] != runtime.base_revision:
        return "REVALIDATE_BASE"
    if cp["policyHash"] != runtime.policy_hash:
        return "REAUTHORIZE"
    if cp["toolSchemaHash"] != runtime.tool_schema_hash:
        return "MIGRATE_INTENT"
    if (cp["model"], cp["effort"]) != (runtime.model, runtime.effort):
        return "REPLAN_FROM_EVIDENCE"
    return "RESUME_PENDING_INTENT"
Enter fullscreen mode Exit fullscreen mode

A model/effort change preserves reconciled effects but invalidates unchanged continuation. Replan from durable evidence, never an opaque transcript.

Resume sequence

Lease c17, reconcile completed effects, compare all revisions, validate evidence, write child c18, then execute its idempotent intent. The lease blocks concurrent resumes; operation-plus-effect identity prevents duplicate writes after a worker dies between tool success and commit.

Failure-injection matrix

Change since checkpoint Required action Forbidden shortcut
none resume pending intent duplicate completed effect
effort only replan from evidence continue opaque partial plan
model revision replan and validate claim semantic equivalence
tool schema migrate or reject intent coerce unknown fields
policy bundle obtain authorization reuse old approval
base commit rerun relevant evidence apply stale patch
unreadable schema quarantine/restart guess missing fields

Normal path: all hashes match, completed effects reconcile, the pending intent is still authorized, and one worker resumes using the stored idempotency key.

Failure path: the tool executed but checkpoint commit failed. On recovery, the effect ledger reports completion; the worker records reconciliation and must not execute again. If the ledger is unavailable, stop rather than infer from model text.

Properties to test

  • P1: each logical effect commits at most once.
  • P2: changed policy requires new authorization.
  • P3: model/effort changes cannot use opaque reasoning as authority.
  • P4: resumed checkpoints have one parent and monotonic sequence.
  • P5: rejection leaves an inspectable decision.

Change one envelope field at a time and inject death around the tool call. Accept only zero duplicate effects and unauthorized resumes.

Tradeoffs are explicit: strict hashes simplify audits but replan more; migrations preserve work but require tests; evidence-only replanning adds calls; restarting reduces compatibility surface while repeating safe computation.

Limits: Anthropic's announcement does not specify this storage protocol or prove compatibility between effort settings. Model identifiers, prices, and setting availability may change. External exactly-once effects require tool cooperation or reconciliation.

Evaluate MonkeyCode with the same invariant

MonkeyCode is a separate candidate: verified material describes an open-source AGPL-3.0 AI development platform, overseas online option, managed server-side cloud development environments, model/task/requirement management, build/test/preview, and free-to-start use. I have not tested checkpoint compatibility there. Use the official campaign link only as an entry to evaluate whether observable evidence can satisfy P1ā€“P5; do not assume internal architecture.

Disclosure: This article promotes MonkeyCode using an official campaign link. Iā€™m a MonkeyCode user, not affiliated with the project, and I receive no commission from this link.

AI assistance disclosure: This article was drafted with AI assistance and reviewed against the cited primary sources.

Top comments (0)