A Model Handoff Also Transfers a State Interface

A mid-run handoff changes more than the model. The receiving agent inherits a working state, some representation of prior reasoning, and any mistakes or assumptions encoded there. That interface can change the result independently of either model's standalone capability.

The signal

The Handoff Tax holds a coding-agent scaffold, prompt, Bash tool, and working tree constant while varying the model switch and transferred trajectory. The study separates escalation to a stronger model from downshift to a cheaper one.

The WTK mechanism signal is a versioned handoff envelope: the state crossing a role or model boundary should be an explicit experimental variable.

The evidence

The primary study covers all 500 SWE-bench Verified tasks, two model pairs, 58 configurations per pair, and 58,000 runs. Under raw full-trajectory transfer, escalation recovered 47 percent of the Claude pair's and 36 percent of the GPT pair's standalone quality gap. Keeping the working tree while dropping the trajectory raised recovery to 64 and 84 percent, but cost more than using the high-capability model alone.

The direction reversed for downshift. Raw transfer retained 50 and 79 percent of the quality gap, while removing prior trajectory reduced retention to 28 and 53 percent. A transcript that burdened escalation helped preserve value when the receiver was weaker.

The boundary

This is external evidence from selected commercial models, one primary coding benchmark, fixed switch points, and then-current prices. The main task-condition cells use one generation per task. Persistent source edits are not equivalent to every WTK artifact, and dropping a transcript may discard audit evidence or undeclared authority along with distracting context.

The builder impact

WTK should qualify the sender, receiver, and exact state envelope together. Durable artifacts, evidence receipts, permissions, unresolved questions, and conversation text are different state classes. A summary must not silently become authority, and suppressed context must remain visible in append-only provenance even when it is not sent to the next model.

The WTK test

On frozen single-agent and team-agent goals, compare a fixed-binding control with predeclared continuation arms that vary only the handoff envelope: full trajectory, bounded typed summary, and durable artifacts without trajectory. Keep package, tools, permissions, targets, models, switch points, evaluator, attempts, and budget fixed. Test escalation and downshift separately.

Measure contract-valid completion, recovered quality, inherited-error rate, missing-evidence failures, authority leakage, cost, latency, and deployment result. Success requires improvement over the current WTK handoff with no loss of evidence or control. Leaked authority, unauditable suppression, or worse deployment is a failure. Mixed directional effects are inconclusive. This is a proposed experiment, not authorization to add dynamic handoffs.

Still unknown

We do not know which state envelope, if any, improves WTK's actual continuation paths without weakening auditability.