Every boundary in agent-assisted work costs the same thing: someone rebuilds context. You hand a task to an agent and write a paragraph of background. The agent hands back forty commits and you rebuild what it was thinking. You come back on Thursday and rebuild what you were thinking on Tuesday. None of that is engineering; it is archaeology.
The fix is not a longer transcript. It is a short, structured, trustworthy record at every boundary. Here is the pattern we use, built on one PatchArc Arc per unit of work.
The unit: one Arc per handoff
Start an Arc when work begins, stop it when it changes hands. The capsule that results is the handoff document: goal, what changed, which scopes, what was verified, the phases, and the decisions and risks recorded along the way. It is computed from evidence, signed, and verifiable, so the receiver does not have to take the sender's word for it.
Human to agent
Before you hand work to an agent, you are usually mid-Arc. Seal what you have and start a new one with the agent's brief as the goal:
$ patcharc stop
✓ Arc sealed: .patcharc/capsules/arc_11aa22.parc
$ patcharc start --detach "Agent: implement the invitation endpoint per arc_11aa22 decisions"Point the agent at the previous review: unzip -p .patcharc/capsules/arc_11aa22.parc evidence/summary/review.json | jq '.decisions, .risks'. That is a few hundred bytes of the exact context it needs, not a transcript to skim.
Agent to human
While the agent works, the observer records its commits. When it is done, you review, and you record what you decide about its work as you go: accepted, rejected, accepted with a risk. Then stop and seal. The capsule now contains the agent's commits and your judgement of them, timestamped, in one signed artifact.
If you use Claude Code, add two lines to CLAUDE.md so the agent records its own design decisions with patcharc decision; they land in the same section, attributed by timestamp.
Today to Thursday
Long Arcs are fine; the observer is a 1.5-second poll. But sealing at the end of a day costs nothing and gives Thursday-you a review to read instead of a diff to reconstruct. Start the next Arc with the open risks from the last one as its goal.
To a reviewer outside the team
Share the capsule or an ArcLink. The reviewer runs patcharc verify (or trusts that the cloud did) and reads the six sections in the order a senior engineer would ask the questions. In 0.2.0 the ArcLink page is minimal and the full review lives in the capsule, so attach the file when the reviewer has the CLI.
Where transcript tools fit
Sometimes the question is not "what was decided" but "what exactly was the agent told". That is a transcript question, and PatchArc does not record transcripts in 0.2.0. Entire stores them on a Git branch with per-commit checkpoints; Lore shares them with a fork prompt; Claude Code has built-in share links. Use one of those for continuity inside the agent loop and a PatchArc capsule for the boundary, and reference one from the other with a patcharc note. The session-sharing guide compares the options.
What to write down, and when
- Decision: any choice that could reasonably have gone the other way, at the moment you make it.
- Risk: anything you noticed and did not fix, at the moment you noticed it.
- Note: test results (until parsing ships), links to transcripts or tickets, anything the next reader needs that is not a decision or a risk.
- Checkpoint: before a step you might want to point back to ("before the schema change").
The honest limits
Test results are not parsed; the transcript is not captured; redaction is not applied at seal; private links have no viewer yet. None of that stops the pattern from working today, because the pattern only needs commits and your one-line entries. It does mean you read before you share.