When the Agent Hands Off to a Human — The Design of *That One Screen*

Agents don't solve everything. At some point they decide I can't or that's outside my authority and hand off to a human. The one screen, one line of that handoff defines the value of the entire automation.

If handoff is designed badly, the human starts the investigation from scratch. In that case the agent didn't move time — it delayed it.

Four handoff patterns we broke

Over nine months, four patterns tried and discarded.

Pattern 1 — raw trace dump

The first. The agent dumps every action it took + every option tried + every external API response — raw. The human sees 47 lines of JSON.

Why broken — reading those 47 lines takes as long as starting fresh. Agent time saved = zero.

Pattern 2 — summary only

Next. The agent hands off a 3-sentence summary. Short, fast.

Why broken — the human doesn't trust the summary. To verify whether the actual cause is in it, they go back to the raw data. They end up reading both — total time increased.

Pattern 3 — decision tree visualization

Next. The agent shows which branches it followed and where it got stuck. Visual.

Why broken — the tree is too big. Agents try dozens of branches. Showing them all is the same as raw dump. Showing only the important branches requires deciding which are important — itself a hard problem.

Pattern 4 — 5W card (the current answer)

What works now. At handoff, a fixed 5W card:

WHAT happened?
  After a deploy at 2026-04-01 14:23 KST, payment success rate dropped -2.3% (SLO -1% breached)

WHEN did the agent stop?
  14:31, after testing three hypotheses inconclusively.

WHY couldn't the agent proceed?
  1. The new code's call to card issuer X behaves differently from other issuers post-merge
  2. That behavior *doesn't reproduce* in our staging environment
  3. *Production rollback authority* is human-only

WHERE to look?
  service: payment-validator-v2
  related deploy: PR#4823, commit a3f7e1
  similar past incident: 2025-11-08 (same issuer, different code path)

WHAT should the human decide?
  Option 1: immediate rollback (5 min, zero impact)
  Option 2: hot fix based on hypothesis 1 (30 min, additional risk)
  Option 3: deeper investigation (gather more triage data)

The card fits in one screen. Humans can decide in 30 seconds.

What makes the 5W card work

The real value of the card is the last line — "What should the human decide?". The agent passes the decision itself to the human, but with the options laid out.

Without that line, humans start investigating again. With it, they only decide.

Handoff accuracy = agent accuracy

After the four-pattern path, our conclusion — the handoff screen's accuracy equals the agent's accuracy. A less smart agent with good handoff lets humans finish quickly. A smart agent with bad handoff — that intelligence doesn't transfer to human time.

An agent's real output is not the action — it's the handoff card.

Generating the card itself

Generating the 5W card is itself an LLM task. At handoff time, a separate LLM call produces the card. That's added cost; it's always net positive against the human investigation time saved.

We tried generating cards via deterministic rules. Readability dropped so much that we abandoned it. LLM-generated card text is the answer.

If an AI agent's human-handoff screen is a raw dump or a summary only, the time the agent saves is evaporating into human investigation time. For teams designing agent-to-human handoff flows right now, building in the 5W card pattern (or a variant) from the start is much cheaper than refactoring later.