Tool Execution
Reliable invocation of MCP tools with timeouts, retries, idempotency, and trace propagation — the boring infrastructure that makes agents trustworthy.
A tool call that times out without idempotency is a duplicate side effect waiting to happen. The execution layer is what makes agent actions safe to retry.
What it solves
Removes the class of bugs where 'the agent did X twice' or 'the agent crashed mid-action' becomes a customer-visible problem.
How we build it
Idempotency keys derived from the workflow context; bounded retries with exponential backoff; explicit timeouts per tool; failure modes (retry, escalate, dead-letter) declared per call site; traces that link every retry to its parent.
- Idempotency keys per invocation
- Bounded retries with backoff
- Per-tool timeout and failure mode
- Trace lineage across retries
What changes when it is in place
Agent retries are safe. Crashes mid-action are recoverable. The runtime stops being the source of incident reports.