Long-running agents now share a standard reliability layer

Claim

Three patterns are now standard. Agent loops run as journaled workflows so completed model and tool calls replay after a crash. Approval is a tool-level flag that pauses the run and persists state until a human answers. Permissions are enforced by OS sandboxes plus a classifier, with deny rules that hold in every mode. Resumption relies on artifacts (git, progress files, feature lists), not conversation memory.

Evidence

• Temporal plus OpenAI Agents SDK GA (2026-03-23): every agent invocation is an Activity; results journaled. Pydantic AI 2.x: journals model and tool calls, "workflows can wait hours or days for a person", engines co-maintained with Temporal, DBOS, Prefect, Restate and Lambda. LangGraph checkpoints after each step; side effects must be idempotent. Restate awakeables and Inngest waitForEvent for approvals.
• Approvals: OpenAI Agents SDK needsApproval with result.interruptions; Vercel needsApproval: true | (input) => boolean with HMAC-signed approvals; LangChain middleware; ADK 2.0 tool confirmation flow (catching BaseException breaks pauses); OpenAI's two triggers are failure thresholds and high-risk irreversible actions.
• Permissions: Claude Code auto mode false-positive 0.4%, false-negative 17% on real overeager actions; six permission modes; deny rules apply in every mode; auto mode drops broad allow rules on entry; a hook "allow" cannot approve critical-path rm. Gemini CLI replaced --allowed-tools with a policy engine; OpenAI Agents SDK 0.17 restricts sandbox materialisation to the base directory.
• Checkpointing (2025-11-26): each session begins with no memory; an initializer writes init.sh, a progress log and a JSON feature list agents may only flip to passing; "compaction alone was not sufficient". Managed Agents now does this persistence server-side.

Open questions

• Temporal and Codex production metrics (talks only). UNVERIFIED.

Sources

• https://temporal.io/blog/announcing-openai-agents-sdk-partnership — blog, Sun Mar 22
• https://pydantic.dev/docs/ai/capabilities/durable_execution/overview/ — docs, Wed Oct 21
• https://restate.dev/blog/build-resilient-ai-agents-with-restate-and-google-adk — blog, Sun Jan 11
• https://developers.openai.com/api/docs/guides/agents/guardrails-approvals — docs, Wed Oct 21
• https://www.anthropic.com/engineering/claude-code-auto-mode — blog, Tue Mar 24
• https://code.claude.com/docs/en/permission-modes — docs, Wed Oct 21
• https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents — blog, Tue Nov 25

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.