Designing the Handoff Contract Between Agents: What Actually Has to Transfer When Work Changes Hands

Splitting work across agents gets most of the design attention: which agent owns which capability, whether a planner-and-executor split beats a single agent with more tools, when the coordination overhead outweighs the benefit. The harder, less-discussed problem sits at the seam between agents — the handoff itself — and that's usually where reliability actually breaks.

What Actually Needs to Cross the Boundary

It's tempting to hand off a natural-language summary: agent A finishes its part, writes a paragraph describing what it did, and passes that paragraph to agent B. This works in demos and fails quietly in production, because a summary is lossy in ways that only show up later. A summary written by agent A, in service of the sub-task A was solving, will emphasize what mattered to A — not necessarily what B needs.

A more reliable handoff carries four distinct things, not one blended narrative: the original user intent in close to its original words, not a re-paraphrase of a re-paraphrase; the constraints A already ruled out, so B doesn't waste a turn re-deriving them or, worse, violate one A had already satisfied; partial results with enough provenance attached — which tool produced them, against what inputs, at what confidence — that B can tell a stale result from a fresh one; and an explicit split between what's settled and what's still open, so B doesn't treat an open question as answered just because it appeared in the handoff text.

Where Handoffs Silently Fail

Three failure modes show up repeatedly once you look for them. First, cross-model lossiness: when A and B are different models (or even the same model at different context lengths), a summary tuned for one model's attention patterns doesn't compress the same way for another, and details that mattered get quietly dropped. Second, no shared error vocabulary: if A's tool call came back empty, B needs to know whether that meant "nothing exists" or "the call failed" — conflating the two is a common source of an agent confidently answering from an empty result set. Third, no attribution trail: when the final output is wrong, an ungoverned handoff makes it genuinely hard to tell whether A handed off bad information or B reasoned poorly from good information — which means every failure turns into a full-stack debugging session instead of a targeted one.

The practical fix looks less like better prompting and more like ordinary API design: treat the handoff as a versioned schema with required fields, not a freeform note. Then write a test that does what you'd do to any other internal contract — deliberately strip or corrupt one field at a time and see which omissions actually change B's output. Whatever breaks first tells you what was load-bearing in the handoff all along, usually well before you'd have found out the hard way in production.