RFC: lambda.execution_timeout — recommendation reversed

> Opened by the knowledge auto-update pipeline. A recommendation **reversed** — nothing was
> rewritten. The maintainer decides the position below; the pipeline then does the typing.

## 1 · What changed

[AWS Lambda durable functions integrates with Pydantic AI](https://aws.amazon.com/about-aws/whats-new/2026/09/aws-lambda-durable-pydantic-ai/)

- was — lambda.md: AWS Lambda execution timeout is a hard 15-minute cap, no durable checkpoint/resume; "eliminates it for minutes-to-hours sessions"
- now — AWS Lambda durable functions (new product/mode) provides checkpointed, resumable execution with fault tolerance across interruptions/timeouts, integrated with Pydantic AI for agent workloads — dimension: standard Lambda (15-min stateless) vs. Lambda durable functions (checkpoint/resume, presumably longer-running agentic workloads)
- still true — Standard AWS Lambda's 15-minute execution timeout and stateless nature remain true for the non-durable Lambda product.

AWS announced "Lambda durable functions," a new execution mode/product that checkpoints progress at each model/tool call, letting a Pydantic AI agent resume after a timeout or interruption instead of restarting from scratch. This is distinct from standard AWS Lambda, which still has a hard 15-minute timeout and no built-in state/resume capability. The old skill recommendation treated Lambda's 15-minute cap as a disqualifier for long-running agent sessions; that blanket disqualification no longer holds for workloads using this new durable mode.

> "AWS Lambda durable functions saves your Pydantic AI agent's progress as it runs, so after an interruption like a timeout, your agent resumes from the last completed step instead of starting over."

## 2 · Recommendations it affects

| location | kind | current text |
| --- | --- | --- |
| `agent-advisor/references/decision-refs/lambda.md:13` | conclusion flips | - Execution timeout: 15 minutes (eliminates it for minutes-to-hours sessions) |
| `agent-advisor/references/decision-refs/temporal.md:54` | derived judgment | - **Lambda classic**: 15-min cap + freeze model is incompatible with a long-poll |
| `agent-advisor/references/decision-refs/temporal.md:96` | derived judgment | \| Short-tool (convert, scrape, validate) \| stateless, seconds, spiky \| **Lambda** (<15 min) |
| `agent-advisor/references/runtimes/lambda.json:11` | derived judgment | "reason": "Lambda has a 15-minute timeout", |
| `agent-advisor/references/runtimes/lambda.json:20` | derived judgment | "reason": "Lambda has a 15-minute timeout", |

## 3 · The decision space

- **Keep the old guidance (Lambda disqualified for minutes-to-hours agent sessions) for standard Lambda, and treat durable functions as a separate, not-yet-vetted option** — depends on: whether the maintainer wants to wait for production track record, pricing clarity, and independent verification before recommending a brand-new GA capability
- **Adopt durable functions as the new recommended path for long-running/agentic Lambda workloads, superseding the 15-min disqualifier** — depends on: confirming GA status (not preview), regional availability matching the workloads in scope, published pricing for checkpoint storage/resume overhead, and at least one real cost/latency benchmark for a multi-hour agent run
- **Split the recommendation: keep standard Lambda's 15-min cap as a hard disqualifier for short bounded tasks where durability adds no value, but allow durable functions as a conditional option for long-running Pydantic-AI-based agents specifically** — depends on: whether the workload is Python + Pydantic AI (the only integration announced) vs. other frameworks/languages, which are not covered
- **Do not change the fact table yet; flag as 'under evaluation' and re-check in the next update cycle** — depends on: maintainer bandwidth and whether any user is currently blocked by the old guidance

## 4 · Proposed position

> Standard AWS Lambda retains a hard 15-minute execution timeout and no native checkpoint/resume — still disqualified for minutes-to-hours stateful sessions run as a single invocation. AWS Lambda durable functions is a separate, newly announced execution mode that checkpoints each step (including model/tool calls) and resumes after interruption, removing the 15-minute wall for supported workloads; treat it as a conditional recommendation, not a blanket replacement: recommend it only for Python agents built on Pydantic AI (the only integration confirmed at announcement), and flag it as new/unproven until GA maturity, checkpoint-storage pricing, and cross-region availability are independently confirmed. Do not cite it yet for non-Python or non-Pydantic-AI agent frameworks, or as a general substitute for standard Lambda in short-task workloads where durability overhead is unnecessary.

**Assumptions to verify before adopting:**

- 'Lambda durable functions' is GA at announcement, not a preview/beta with different SLAs
- Regional availability of durable functions matches the regions the skill's target workloads actually run in (announcement claims 'all Regions where Lambda durable functions is available' but does not enumerate them)
- Checkpoint/resume pricing (storage of intermediate state, per-step overhead) is publicly documented and not materially higher than re-running short tasks
- The integration is currently scoped to Python + Pydantic AI only; other languages/frameworks are unsupported and this fact should not be generalized
- No independent benchmark yet exists confirming resume behavior avoids duplicate side effects (e.g., double billing) under real failure conditions, despite the vendor's claim
- Cold-start and checkpoint latency overhead versus standard Lambda has not been measured and could affect cost/latency-sensitive workloads
- The 'minutes-to-hours' session claim assumes no other AWS service limit (e.g., API Gateway, event source timeouts) truncates the session upstream of Lambda itself
- This is a vendor-authored announcement; maturity claims should be treated skeptically until third-party validation or maintainer's own testing confirms behavior under failure/retry scenarios

## Decision — tick one; the next run acts on it

- [ ] **Adopt the proposed position** — the pipeline rewrites the affected locations and opens a draft PR for review
- [ ] **Adopt with changes** — edit the "Proposed position" text above first, then tick this
- [ ] **Reject** — close this issue; nothing is rewritten

Read the proposed position and inspect agent-advisor/references/decision-refs/lambda.md, agent-advisor/references/decision-refs/temporal.md, and agent-advisor/references/runtimes/lambda.json. First verify the listed assumptions about durable functions' status, regions, pricing, integration scope, and behavior; done requires a maintainer decision and consistent updates to the affected references through the pipeline.

Written by the indexing model from the issue text.