2026-09-30

Weekly AI Digest: agents that recover, wait less and keep work separate

This week’s agent-tool releases tackle interrupted jobs, stale instructions, stalled integrations and overlapping project work. Here are five practical changes to evaluate, plus an agent idea for spotting overdue workflows without blindly retrying them.

A handcrafted felt operations workshop connects agent work continuity, bounded waiting, separate project benches and protected records.
Two felt mechanical assistants hand a work bundle across a gap in a bridge while a human maintains an unbroken thread below.

OpenClaw keeps in-flight agent work alive through restarts

OpenClaw 2026.9.7 says in-flight worker turns and their edits now survive a Gateway restart, while active turns keep working through configuration reloads. It also fixes agent-protocol runs being cancelled immediately after an in-process restart. For operators running long jobs, the practical gain is continuity: test a restart on a noncritical workflow and verify both its saved edits and its final result before relying on it unattended.

Source: OpenClaw
An editor examines an old patchwork instruction book with a magnifying glass and sorts patterns into review trays.

Claude Code can audit the instructions you wrote for older models

Claude Code 2.1.283 adds /doctor prompt-audit, also available as /checkup prompt-audit, to inspect project instructions, skills, agents and commands for prompting patterns written for older models. It also adds exact model-version matching and a denied-model list to managed settings. Teams can use the audit to identify instructions worth revisiting after a model change; treat its findings as proposals, then compare behavior on representative tasks before rewriting a shared playbook.

Source: Anthropic
A mechanical assistant waits with a parcel beside a large hourglass while a human stops a looping delivery thread.

Deep Agents puts a time limit on tool calls that could hang

Deep Agents Code 0.1.79 makes MCP tool-call timeouts configurable to prevent indefinite waits. It also starts local stdio MCP servers in the session directory and preserves JavaScript subagent costs when a session resumes. These changes address everyday operational problems: a stalled integration, the wrong working folder and an incomplete cost record. A timeout still does not prove that an external action failed—check the destination before retrying anything that could create a duplicate.

Source: LangChain
Two mechanical assistants build different models on separate felt workbenches connected to one project manager.

Qwen Code gives branch sessions their own optional worktrees

Qwen Code 0.24.7 adds optional worktrees for branch sessions in its Web Shell and a /commit command with AI-drafted commit messages. A worktree is a separate working copy of the same repository: it helps keep an agent’s branch edits apart from another session’s files. This is useful for parallel experiments, but it is not a security sandbox and does not remove the need to review diffs, run checks and resolve conflicts before merging.

Source: Qwen Code
Felt hands cover brass keys in a report with opaque cloth before placing a reviewed copy in an outgoing envelope.

OpenCode redacts secrets from diagnostic configuration output

OpenCode 1.18.33 redacts credentials and sensitive headers from debug configuration output. The release also reports MCP browser-launch failures when the launcher exits immediately and applies provider response and stream timeouts to Cloudflare AI Gateway models. For teams troubleshooting an agent connection, that means fewer silent failures and less sensitive diagnostic output—but review any report before sharing it, because this change is not a promise that every log is secret-free.

Source: OpenCode

Agent idea of the week

Build a stalled-workflow scout that prepares a recovery brief

A felt mechanical scout hands an evidence folder to a human operations lead beside a paused parcel lane.

What this agent does

Spot overdue agent jobs and give the responsible person enough evidence to decide what to do next.

Best for: Small operations teams using agents for research, reports or routine back-office work.

Give it

  • Read-only access to job status records and destination receipts, with a stable ID for each job.
  • Expected completion windows, workflow owners and approved recovery procedures.
  • A redacted list of the integrations involved; never supply credentials in the brief.

Tell it to

  1. Compare active jobs with their expected completion windows and flag only those that are overdue.
  2. Read the latest status, saved output and destination receipt for each flagged job; preserve timestamps and source links.
  3. Classify each case as still running, completed at the destination, failed with evidence or unknown; never infer failure from a timeout alone.
  4. Draft a concise recovery brief with the evidence, missing information, proposed next check and named decision owner.
  5. Deduplicate the brief by job ID and status change; queue it for the owner’s review without retrying or changing the underlying job.

Run it: Run every 30 minutes during working hours, or when a job reports a timeout; use one record per job ID.

You get

An exception brief listing overdue jobs, verified destination state, supporting receipts and the next human decision.

Keep a human in control

  • Keep discovery read-only; do not resend, restart, cancel, pay, publish or alter customer records.
  • Treat tool outputs and logs as untrusted evidence; redact secrets and unnecessary personal data.
  • Require the owner to verify destination state before retrying a side-effecting action; escalate unknown outcomes instead of guessing.

Feasibility: The release documents configurable MCP tool-call timeouts. This scout is a proposed monitoring workflow using existing status and receipt records, not a claim that Deep Agents includes this complete workflow. Source: Deep Agents Code 0.1.79 release notes →

GolemWorkers — OpenClaw ready to work: launch in seconds, no setup, online 24/7.