How can a long-running agent resume safely after hours or days?

Scope

Use this pattern for agents that pause for humans, rate limits, scheduled time, asynchronous jobs, external events, or work spanning multiple model sessions. It focuses on durable execution state. Commit-time authorization expiry and duplicate side effects have separate guides because they require additional controls.

Why it happens

A paused conversation looks like state: it contains the plan, tool results, and a final sentence such as “waiting for approval.” Hours later, however, the world has changed. A deployment moved, data was edited, a token expired, another worker completed the action, or code now interprets the transcript differently.

Model context is also lossy. Compaction may omit an exception, and regenerating a summary can change meaning. Safe resumption requires a system of record that names completed transitions and external effects independently of prose.

Diagnosis

Crash the workflow at four points: before an external call, after the call but before recording its result, after recording success, and while waiting for approval. Restart with a different worker and, if supported, a newer code version.

For each run, ask:

• Can the engine identify the exact last committed transition?
• Can it distinguish not attempted, in flight, committed, failed, and outcome unknown?
• Are artifacts immutable or version-addressed?
• Can it detect that target state, policy, identity, approval, or deadline changed?
• Does replay repeat a side effect?
• Can an operator explain and override the recovery decision?

Solution

Persist a checkpoint after deterministic state transitions and before exposing the next action. Record external calls in an effect ledger with idempotency keys and target references. When a crash leaves the outcome unknown, query the target by that key or reconcile observed state; never assume failure and resend blindly.

On wake, verify checkpoint integrity and code compatibility. Refresh current resource state, policy, identity, approval, deadlines, and dependency versions. Rebuild model context from the typed snapshot and authoritative artifacts. The next transition should evaluate explicit preconditions and create a new decision record.

Migration deserves a first-class path. If old histories cannot replay under new code, route them through a reviewed migration or complete them with the old worker version.

Artifact

The resume contract contains:

Contract area	Required fields
Workflow identity	Task, tenant, workflow type/version, run, owner, and creation time
Checkpoint	Sequence, integrity hash, last transition, state schema, and code version
Artifacts	Stable IDs, versions, provenance, and access policy
Effects	Proposed/in-flight/committed/failed/unknown state, target, and idempotency key
Wait condition	Event, deadline, lease, approver, and cancellation rule
Authority snapshot	Actor, scopes, policy, approval, target, and expiry
Wake checks	Current state, authority, deadlines, dependencies, and conflict detection
Next transition	Preconditions, allowed tools, budgets, and rollback/compensation
Handoff	Owner and packet when automatic reconciliation fails

Common mistakes

• Treating a chat summary as the workflow database.
• Replaying nondeterministic code without version control.
• Retrying an external action whose previous outcome is unknown.
• Resuming with the identity and approval captured days earlier.
• Migrating state schemas without testing historical checkpoints and effect records.