Handoffs

Most complex tasks have a nuanced final tenth that needs human judgment. The trick isn't full automation — it's a clean handoff at the right moment, carrying just enough context to continue.

Why this, for you: the choice between cold-starting from raw output and over-automating the tricky tail is a false one. A structured handoff gives you the agent's work plus its reasoning at the exact point where your judgment is worth more than its persistence.

Background agents handle well-defined work efficiently, but the final ~10% often requires judgment, taste, or infrastructure the agent can't reach. The background-to-foreground handoff transfers ownership at a planned threshold — not at failure, not at 100%.

1 Stop at the judgment threshold, hand off four things

The ~90% figure is illustrative; the real criterion is "the remaining work requires human judgment" — confidence falls below threshold on an externally-impactful decision, multiple valid options reflect preference not correctness, or testing needs infrastructure the agent can't reach. Four components make the handoff work:

Completion threshold — the agent stops before nuanced judgment calls.
Distilled summary — a compact note (what's done, what remains, decisions made and alternatives rejected, open questions), not the full trace.
Artifact handoff — a draft PR on a named branch plus a progress file (claude-progress.txt) and git history; the human reads diffs, not agent logs.
Tool parity — the human picks up in the same IDE, terminal, and MCP servers, so the summary applies directly.

Without a reliable stopping signal, the handoff fails in the worst way: the agent stops too late, after it has already made the irreversible judgment calls the handoff existed to preserve for you. Progress visibility is the prerequisite — you need a signal to know when to engage.

2 Three handoffs, three different jobs

This pattern is often conflated with two neighbors. They differ by trigger and direction:

Handoff	Trigger	What moves
Human-in-the-loop gate	Before an irreversible action	Agent pauses; you approve
Cloud–local handoff	A surface transition	Same work, different infra
Background→foreground	Near-complete task	Ownership: agent → human

And the in-thread side-channel is a fourth move in the other direction: a tagged sub-conversation lets you ask a mid-task clarifying question inside one session, then returns the agent to its prior goal — without spawning a new session or losing context.

When a hard 100% loop is safer

The handoff assumes the agent can identify the threshold and write an accurate summary. Both can fail. Summary drift — a note that omits or misrepresents decisions — drops you onto a false starting point, and a misleading draft PR can be harder to untangle than starting fresh from requirements. Tool parity absent (different branch state, missing MCP servers, inaccessible infrastructure) degrades the handoff to a cold start regardless of summary quality. In those conditions, full automation with post-hoc review is often more reliable.

↪ Your win: take over at the right moment, with context intact

Stop at the judgment threshold — not at failure, not at 100%; hand off the nuanced tail.
Pass a distilled summary — done, remaining, decisions, open questions — not the raw trace.
Hand off durable artifacts — draft PR + progress file + git history you can pick up from alone.
Keep tool parity — same IDE and servers, or the handoff degrades to a cold start.
Know the four directions — gate, cloud–local, background→foreground, in-thread side-channel.

Retrieval practice — recall, don't peek

Question 1The background-to-foreground handoff fires when the task is…

Question 2The handoff passes the human…

Question 3The worst stopping failure is when the agent stops…

Question 4Missing tool parity degrades the handoff into…

Question 5 · spaced recall from Lesson 6In a headless CI run, you must set --max-turns because…

Ask me anything. Want a claude-progress.txt template that captures decisions and open questions, or how the in-thread side-channel returns an agent to its prior goal after a clarification? Next, Part 3 opens with Snapshot and Roll Back — reversible repo setup.

Handoffs

1 Stop at the judgment threshold, hand off four things

2 Three handoffs, three different jobs

When a hard 100% loop is safer

↪ Your win: take over at the right moment, with context intact

Go deeper