A fresh agent shouldn’t have to rediscover the project every time.

I heard that someone was using an Obsidian vault to make fast handoffs between agents, and brought it up with Claude Opus 5. Opus questioned whether that approach solved the deeper problem. I kept probing: did a newly spawned agent begin with the previous agent’s knowledge? The answer was no. That was the real cold-start problem, and the conversation grew into WarmStart.

A vault is good at preserving project information. WarmStart does a different job: it compiles selected, current, evidence-backed context into a reviewed startup packet for the next agent.

The design in one line: preserve the history, then prepare the few current facts the next agent needs.

A vault could be a future input to WarmStart. No such integration exists today.

A new agent session often begins cold. It re-reads the same files, re-asks the same questions, and re-learns the same project rules before its first useful action. Sometimes it starts from context that has quietly gone stale, which can be worse than starting from nothing.

This is the same family of problem as compaction, where an agent loses what it knew in the middle of a job. WarmStart looks at the other end: the first minutes of a brand-new session.

1. Generate

   The tool reads the project and prepares the notes on your computer, without calling an AI model.

2. Review

   Notes pass through a review step, and an approval is recorded before handoff.

3. Publish the notes

   Approved notes go into a marked section of the project’s AGENTS.md. A verified connection that hands them to an arriving agent is still missing.

4. Watch for staleness

   Context can drift out of date. Stale-context failures are one of the things the matched trial will count.

The repository has the code and full instructions for generating, approving and updating the notes.

• Review before use. Publishing needs a recorded approval. In the saved demonstrations, an agent made the test approvals, so a record alone doesn’t show that a person read the notes.
• Approval is a separate step. Generating context does not hand it to an agent by itself.
• Staleness is a known failure mode. Stale context is tracked as its own kind of failure.
• Privacy is checked. Privacy and wheel-exclusion checks are part of the release gate.

Benefit: inconclusive Release checks: passed Recovery failures reproduced

Trying the tool on ten repositories exposed four defects in its own checks. Those findings helped us improve the release. The next question is whether the notes help an agent do its job.

October 9: testing interrupted and overlapping writes

We ran six selected publication schedules in disposable repositories, including interrupted writes, two publishers at once and edits made while a publisher was paused. Three cases still failed after an ordinary recovery: missing publication records, lost newer approved notes and overwritten human edits.

Inspecting, compiling and checking stale file dependencies worked in this batch. Fresh-agent delivery was not exercised: no arriving learner was launched, and no automatic connector was verified. The next steps are publication repair and a real delivery check before measuring whether the notes help.

• The agent received no WarmStart notes, so the trial didn’t test the tool’s intended benefit.
• We also needed a second run of the same task without the notes, for comparison.
• Speed, cost and quality benefits have not been measured yet.
• Compatibility beyond the pilot has not been shown for any model family.

Give agents the same task. Some get the checked project notes; others don’t. Repeat the comparison on different models.

We’ll publish both the benefits and the pain points. If WarmStart doesn’t help, that result goes on this page too.

This is an open invitation to test, not a claim of compatibility. These additional model families haven’t been tried yet.