Engineering / research
Durable execution for tool-using agents
Model agent runs as recoverable workflows with explicit tool contracts, budgets, approvals, cancellation, and observable side-effect boundaries.
Your route through engineering
Follow the path or begin with a question. Every lesson is open.
Your place, at your pace. Progress is optional and saved only in this browser when available.
Read / explore / reason
Inside this module
Module 01 research
Durable execution for tool-using agents
Model agent runs as recoverable workflows with explicit tool contracts, budgets, approvals, cancellation, and observable side-effect boundaries.
- 01
Run state machines: separate logical intent from a worker incarnation
Model durable run transitions, revision checks, stale completions, and external-effect uncertainty under an explicitly proposed runtime protocol.
27 min - 02
Typed tool contracts: validate inputs without assuming atomic execution
Separate schema validation, domain constraints, authorization, resource limits, output validation, and partial side-effect reporting.
27 min - 03
Step journals: replay recorded decisions without rerunning the outside world
Define immutable event records, artifact references, deterministic control replay, and recovery for unresolved tool intents.
28 min - 04
Tool idempotency: preserve the effect identity across retries
Separate transport attempts from logical operations, define payload-bound deduplication, and account for provider retention and unknown outcomes.
28 min - 05
Effect commit and compensation: recover without pretending the world rewinds
Separate local commit boundaries, external receipts, transactional outboxes, and conditional saga compensations.
28 min - 06
Scopes and approvals: bind authority to the exact reviewed effect
Separate credentials, scoped capabilities, approval receipts, artifact identity, and revocation timing in a proposed execution gate.
28 min - 07
Concurrent tools: schedule the dependency graph, not the visible list
Separate data dependencies, resource conflicts, deterministic joins, structured child ownership, and bounded parallel execution.
28 min - 08
Cancellation: distinguish the request from quiescence and cleanup
Define nested deadlines, cooperative cancellation, child ownership, external-effect reconciliation, and bounded cleanup responsibilities.
28 min - 09
Runtime budgets: reserve bounded work before dispatching it
Separate usage estimates from enforceable reservations across tokens, tool attempts, compute time, retries, and concurrent children.
28 min - 10
Recovery traces: bind the observed run to artifacts and fault oracles
Connect causal tracing with durable identities, reviewed artifacts, authoritative effect counts, and crash-window tests.
29 min