# Run a long or scheduled agent job that survives restarts

> Run a long or scheduled agent job that survives restarts: 0 companies fully do it, 18 with a limit, 25 carry the tag. Every listing in Agent frameworks & SDKs with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/frameworks.durable-long-run/
- Markdown: https://www.anchorterminal.com/niches/frameworks.durable-long-run/index.md (~2,000 tokens)
- Slim: https://www.anchorterminal.com/niches/frameworks.durable-long-run/index.min.md (~430 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/frameworks.durable-long-run/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

No company's docs show a full fit yet. The listings below do the job with a limit, or not at all.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 0 | 18 | 25 | 11 | 9 |

## What counts as done

State is written to a named durable store after each step, a run killed mid-way resumes from the last completed step with a documented rule for the tool call that was in flight, cancelling keeps what was already streamed, and the docs show resume across a code change or new agent ids.

Measured, for our own tests: kill the process during step 3 of 5 and count repeated tool calls on resume.

How agents ask for it: run this report for every customer each night; resume the agent where it stopped after a crash; keep the conversation after the server restarts; cancel a run and carry on the thread

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **AgentOS (Framers Lab, Inc.)**: SqlCheckpointStore (SQLite or Postgres) ships, and resume marks write and external nodes complete with their recorded output while failed nodes re-run. The docs don't cover resume across code changes or what a cancel keeps.
- **Agno (Agno Inc.)**: Runs persist to SQLite or Postgres and continue by run_id in a new process. I found no rule for a tool call in flight at a crash, and resume across code changes isn't documented.
- **BoxLang AI (Ortus Solutions, Corp.)**: A checkpointer memory (cache or jdbc) stores suspended state and cancelRun stops at the next beforeLLMCall or beforeToolCall. No documented rule for a tool in flight at a crash or resume across code changes.
- **Claude Agent SDK (Anthropic, PBC)**: Sessions persist and resume by id, and a SessionStore mirrors transcripts to your own store so a new container can hydrate. I found no rule for a tool call in flight when the process died; state on the container disk is lost on restart without a store.
- **CrewAI (CrewAI, Inc.)**: Checkpointing ships JsonProvider and SqliteProvider; a restore skips completed tasks and rehydrates memory. No Postgres or Redis store, no rule for an in-flight tool call, and lambdas and bound-method callbacks are not restored.
- **Cursor Agent SDK (Anysphere, Inc.)**: Sessions keep an append-only record of messages and tool calls that survives a restart, stop records turn.failed as cancelled, and tool effects mark writes run-once. The store is not named and resume across code changes isn't covered.
- **Docker Agent (Docker, Inc.)**: Every message, tool call and sub-agent run is stored in a SQLite session.db and a run resumes by session id. No rule for a tool call in flight at a crash, and resuming mid-run differs from continuing a conversation.
- **Agent Development Kit (ADK) (Google LLC)**: ResumabilityConfig logs completed steps and reinstates tool results, re-running the interrupted tool so tools run at least once; resume after modifying the workflow is not supported, and resume isn't available from ADK Web or the CLI.
- **guilds.run (Cigale AI)**: Runs live in PostgreSQL; on restart, unfinished committed tool calls get an interrupted result, pending group dispatch resumes, and routines wake agents on schedules. Resume across a code change isn't documented, and a run that hit its tool-call cap can't be resumed.
- **LangGraph (LangChain, Inc.)**: Postgres, SQLite and MongoDB checkpointers ship; checkpoints land at super-step boundaries and resume re-runs the node at least once, with idempotency-key guidance. Accounts: cancelling a run loses state since the last checkpoint (langgraph #5672, open) and Redis resume failed (#5074).
- **LlamaIndex Framework (LlamaIndex, Inc.)**: Context.to_dict() snapshots a run for any store, resume re-dispatches pending events and re-runs a step that was mid-execution (at-least-once), and a DBOS integration exists. You write the store; an account hit serialisation errors for Gemini function calls in Postgres (llama_index #19992).
- **Mastra (Kepler Software, Inc.)**: Durable agents run the loop inside a workflow; with Inngest each tool call is a memoized step retried independently. Storage is libSQL, PostgreSQL or MongoDB. Agent run snapshots are deleted when the run finishes, and resume across code change isn't covered.
- **Microsoft Agent Framework (Microsoft Corporation)**: Workflow checkpoints at superstep ends go to in-memory, file or Cosmos DB storage, and the Durable Task extension keeps agent sessions across crashes for days. Upgraded workers need replay-compatibility checks, and agent-id changes broke resume for one account (#4793).
- **OpenAI Agents SDK (OpenAI OpCo, LLC)**: Sessions ship for SQLite, SQLAlchemy, Redis and Dapr, and a RunState serialises with to_string() to resume an interrupted or cancelled run. No rule for a tool call in flight at a crash was found.
- **Pydantic AI (Pydantic Services Inc.)**: Nine durable engines (Temporal, DBOS, Prefect, Restate, AWS Lambda and more) journal model and tool calls so a run survives a crash mid-step, and StepPersistence saves, resumes and forks runs. The overview doesn't cover resume across code changes or what a cancel keeps.
- **Strands Agents (Amazon Web Services, Inc.)**: Session managers persist to local disk or S3 (SnapshotSessionManager recommended), with snapshot_latest used to resume after a restart. No Postgres or Redis store and no rule for a tool call in flight; the in-memory history gap of strands #33 is closed.
- **TradingAgents (Tauric Research)**: Opt-in --checkpoint saves LangGraph state after each node to a per-ticker SQLite database so a crashed run resumes from the last step. No rule for an in-flight tool call; its tools only read data, so repeats have no side effect.
- **AI SDK (Vercel Inc.)**: WorkflowAgent from @ai-sdk/workflow runs each tool call as a durable workflow step with retries and survives restarts; ToolLoopAgent itself is in memory. Durability depends on Vercel Workflow, and WorkflowAgent applies no default step limit.

- not this job: NitroStack (Nitrostack Inc.): MCP Tasks cover long-running tool calls with progress, cancellation and a TTL, but there is no agent run state or checkpoint store; an account saw tasks still working past their TTL deleted (nitrostack #329, open).
- not this job: smolagents (Hugging Face, Inc.): No checkpoint store or resume page in the docs index; an open account proposes state serialisation because none exists (smolagents #364).

## What stops agents finishing

- only an in-memory or file checkpoint store ships
- history with provider-specific function-call objects fails to serialise to the store
- resume fails when agent ids or code change between runs
- a cancelled run loses everything after the last checkpoint

## Around it

Before: agent-harnesses.scheduled-job. After: human-in-the-loop.durable-wait, agent-observability.debug-failed-run.

## Where this comes from

Weight 11 of 100 in Agent frameworks & SDKs: 7 of 61 accounts have this as their main job (share 11.5, rounded by largest remainder to sum 100)

As data: https://www.anchorterminal.com/api/v1/use-cases/frameworks.durable-long-run.json, or the MCP tool get_use_case with id frameworks.durable-long-run. Draft, read 2026-11-15 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
