Context Engineering Handoff Beats

Accepted Beat Path

1. Context Has Half-Life: a long run accumulates old guesses, failed commands, partial plans, stale reads, rejected paths, and recency residue.
2. Rot vs Poison: rot is expired truth; poison is wrong material that keeps steering the run.
3. The Cost Shows Up Before the Limit: token pressure creates latency, spend, and attention drag before the window is full.
4. Curated Continuation: the key question becomes what should still shape the next decision.
5. What the Skill Enforces: handoff creates a temporary continuation artifact with suggested skills, durable references, and redaction.
6. When to Hand Off: hand off before a fresh, long, consuming phase; preserve flow inside a phase.
7. Bad Handoff vs Good Handoff: bad copies transcript residue or interrupts working memory; good carries only the minimum state for a clean restart.

Beat Drafts

1. Context Has Half-Life

Start from the thing every long agent run eventually makes visible: context decays. The run begins with clean instructions, useful evidence, and a tractable objective. Then the transcript starts collecting residue: old guesses, failed commands, partial plans, stale file reads, rejected paths, and whatever happened most recently.

The point of this beat is to make context feel active rather than inert. A long context window behaves like a working environment where older material can keep affecting newer decisions. That is why Harness treats handoff as part of context engineering, not as a clerical wrap-up.

2. Rot vs Poison

Separate the two ways context goes bad.

Context rot is expired truth. The agent saw something useful earlier, but the run moved on. A test failure was fixed. A plan was replaced. A file moved. A root-level instruction was narrowed by a scoped instruction. The old fact has passed its sell-by date.

Context poisoning is different. It is wrong material that keeps exerting force because it remains visible. A bad hypothesis, a noisy log excerpt, a hallucinated constraint, or an abandoned implementation path can keep nudging the model back toward work that should be dead.

This beat gives the reader vocabulary. Rot asks for pruning. Poison asks for quarantine. Both are reasons to stop trusting raw transcript continuity as the default form of memory.

3. The Cost Shows Up Before the Limit

Move from correctness risk to operating cost. A context window does not need to be full before it starts changing the quality of the run.

Every extra chunk of old material adds work. The model has to separate current signal from stale residue. The user pays for more tokens. The turn takes longer. The agent has more chances to anchor on old details because those details are still present and linguistically available.

This is the practical version of context management. The question is wider than "will the next message fit?" The run can become slower, more expensive, and less precise while it still technically fits.

This beat should make manual handoff feel like an efficiency tool, not only a safety tool. A clean restart can reduce noise before the run collapses under its own transcript.

4. Curated Continuation

Pivot from window size to authority. The question stops being "how much context can we keep?" and becomes "what should still be allowed to shape the next decision?"

That shift is the core of manual handoff. The agent selects the pieces that still matter: the current objective, accepted decisions, changed files, verification evidence, unresolved questions, and the skills the next run should activate.

This is where handoff becomes a context-engineering move. It turns continuity from accident into design. The next session starts from curated state instead of inherited residue.

This beat should also set up the contrast with automatic compaction. Compaction summarizes what happened. Handoff declares what remains operational.

5. What the Skill Enforces

Make the handoff concrete by naming the contract.

The handoff skill asks for a continuation document rather than a preserved chat. The document goes to the operating system's temporary directory, outside the current workspace, so it behaves like a transfer artifact rather than another project file.

It must include suggested skills. That matters because the next session should not rediscover the workflow from scratch. If the continuation needs docs-ingest-phase, writing-shape, review-phase, or a scoped implementation skill, the handoff should say so directly.

It must reference durable artifacts instead of duplicating them. Specs, plans, ADRs, issues, commits, and diffs already have homes. The handoff points to them by path or URL and carries only the connective tissue the next agent needs.

It must redact sensitive information. A handoff is designed to travel across sessions and agents, so secrets, credentials, API keys, passwords, and personal information do not belong there.

The result is small on purpose. A good handoff carries the minimum state required to restart clean.

6. When to Hand Off

Make the operational rule explicit: handoff is for starting fresh with proper context before a new long phase.

Good moments are phase boundaries. A requirements pass is done and planning is next. A plan is accepted and implementation will be long. Debugging produced enough evidence that resolution should start in a cleaner run. A docs pass is about to turn scattered implementation reality into durable wiki pages. The current session has done its job, and the next one needs a clean entrypoint.

The bad moment is inside the same phase while the agent still has live working state. Interrupting implementation halfway through a coherent edit, stopping a debugging hypothesis before it has been verified, or breaking a writing pass while the shape is still forming can destroy useful continuity. That move is self-sabotage with better stationery.

So the rule is: hand off when the next run should start fresh, not when the current run is merely uncomfortable. Preserve flow inside a phase. Reset between phases.

7. Bad Handoff vs Good Handoff

Use contrast to make the standard memorable.

A bad handoff copies the transcript. It carries every false start, every noisy command, every abandoned theory, and every accidental detail into the next run. It feels thorough, but it recreates the same context pressure in a smaller file.

A worse handoff happens too early. It interrupts the same phase, forcing the next agent to reconstruct working memory that was still useful: why this edit is half-finished, which branch of the hypothesis tree is currently active, what local assumption has not yet been tested.

A good handoff is selective. It says: here is the objective now, here is what changed, here is what was verified, here is what remains open, here are the artifacts to read, here are the skills to activate, and here is what not to repeat.

The goal is a next session that begins cleanly without becoming naive.

Parked Beats

• A future lifecycle-flow page could show the exact shape of a handoff document produced at the end of a delivery phase.

Gaps

• The current fragments do not include a fully worked handoff example. Add one only if the final page needs more teaching surface than the current concept-page section.