Product
Retry Policies
Per-tool exponential backoff, declared in one file instead of scattered through your handlers.
Retry logic written inline drifts: every handler grows its own variant, and none of them agree on how many attempts is too many. A policy file makes the rule explicit, reviewable, and identical everywhere it applies.
Per-tool overrides
A flaky search endpoint can retry three times while a refund retries once, from the same file.
Validated at deploy
Rules that can never be reached are rejected before they ship.
Budgets, not just counts
A policy can cap total spend on retries as well as total attempts.
retry: {
"kb.search": { max: 3, backoff: "exponential" },
"stripe.refund": { max: 1, requireApproval: true }
}Measured outcome
A team replaced four thousand lines of hand-written backoff across nineteen handlers with one policy file and deleted the rest.
Product
- Run GraphWatch every step, branch, and retry as it happens, with the exact inputs and outputs on each node.
- Audit TrailEvery prompt, tool call, and token — retained and replayable, exportable as compliance evidence.
- Human in the LoopPause a run, review the state, approve or amend, and resume from the same step.