Engineering Notes / 48h Reliability Sprint Packet
48h Agent Reliability Sprint: packet
A concrete, minimal “what you get” doc you can link in outreach without relying on GitHub/X.
When this is a fit
- Your agent works in demos but flakes in production (timeouts, partial state, tool failures, retries spiraling).
- You can’t reliably reproduce failures from a single trace/log bundle.
- You need a fix path that engineers can implement, not generic advice.
What we deliver in 48 hours
- Repro harness: a minimal script / test that triggers the failure deterministically (or a hard proof it can’t be reproduced with given inputs).
- Root-cause analysis: where and why the stack loses invariants (state drift, tool I/O, retries, rate limits, timeouts, model/tool mismatch).
- Fix path: specific code-level changes, with acceptance criteria (what must be true after the fix).
- Reliability guardrails: timeouts + retries + circuit breakers + idempotency + run correlation, sized to your workload.
Guarantee: if we can’t produce a reproducible test + fix path in 48h, we don’t invoice.
What to send us (minimum)
- One failing run trace (logs + prompts/messages + tool calls + tool outputs).
- Environment + versions (models, SDKs, orchestrator, tool APIs, timeouts/retry policy).
- A crisp success criterion (e.g., “must never double-charge”, “must always call tool X before tool Y”, “latency < 4s p95”).
If you can’t share full data, we can do a redacted bundle or work against synthetic traces.
← Back to service page
