@hanakoxbt: your two agents are not collaborating. the second one is redoing the first one's work. it looks like teamwork in the tr…

your two agents are not collaborating. the second one is redoing the first one's work. it looks like teamwork in the trace. two clean handoffs, one answer. what actually happened is that agent A paid for six findings and agent B inherited one sentence. two agents cannot share a window. that is not a framework limitation, it is the whole mechanism. so the first one summarises, and the second one inherits the summary. A read the docs, sent a null field, got a 400, tried the batch route, got a 404, and worked out that the endpoint wants ISO dates and an auth header. five of those six findings cost a tool call. what crosses is one line. use v2, ISO dates, no nulls. the 400 does not cross. neither does the request that caused it, the response body, or the routes already ruled out. > why this is worse than a lossy summary it is not that B distrusts the line. it believes it completely. it is that a conclusion without its evidence cannot be reasoned from. B has a sentence saying nulls are rejected. it does not have the failure that proves it, so the first time it hits an edge that sentence does not cover, it has nothing to work with. and an agent with a gap does what any agent does. it goes and finds out. finding out means the same doc read, the same null field, the same 400. by the time B arrives where A already was, that stretch has been paid for twice. > what to actually do pass artifacts, not prose. a file, a schema, a captured response. those survive the handoff intact, because they are not a description of evidence, they are the evidence. write down what failed, not just what worked. the ruled-out paths are the part that stops the next agent walking them. check who the summary is addressed to. one written for a human reads as a report. one written for the next agent reads as a spec, and they are different documents. and before you split the work at all, ask what the second agent will need to have seen. if the answer is most of what the first one saw, you do not have two agents. you have one agent and an expensive amnesia. this is also why measuring the final answer tells you nothing here. both runs end correct. the duplicated work is only visible in the trajectory. save this - then read the eval setup below