A research assistant runs a planner that fans out to four worker agents and a synthesiser that writes the final answer. One worker retrieves the wrong document and returns a fluent confident summary of it. What happens downstream and what stops it?

Step by step, what happens

1. The worker returns prose. Nothing errors: the tool call succeeded, the model produced well-formed output, every latency and error-rate metric is green.
2. The synthesiser receives four summaries with no way to distinguish them. It has not seen the source documents, so it cannot check the claim, and it has no signal that one input is less trustworthy than the others.
3. The synthesiser does what it is built to do: it reconciles the inputs into a coherent narrative. A confident wrong summary is easier to integrate than a hedged correct one, so the bad input is often weighted more heavily.
4. The final answer carries the error with the authority of a four-source synthesis, which readers trust more than a single-source answer.

Where it amplifies

The arithmetic is unforgiving. If each of six steps is independently correct 97% of the time, the pipeline is correct about 83% of the time (0.97^6). Adding a seventh specialised agent to improve one step by two points can easily lose more than that to the extra handoff. This is why multi-agent decompositions frequently score worse end to end than a single agent with the same tools, despite each component testing well.

Prose handoffs are the specific amplifier. Every handoff that passes a summary rather than an identifier destroys the ability to check the claim later. The synthesiser cannot verify what it cannot trace.

What stops it

• Pass references rather than prose. A worker returns claims paired with document ids and character offsets. The synthesiser is then able to quote and, more usefully, a deterministic checker can confirm that every sentence in the final answer maps to a retrieved span.
• A verification step that reads the source, not the summary. One extra call that checks the top claims against the retrieved text catches exactly this failure, because the error is invisible at the summary layer and obvious at the source layer.
• Force disagreement to surface. If two workers cover overlapping ground, have the synthesiser report the conflict rather than smooth it. Silent reconciliation is where this failure becomes invisible.
• Abstention as a first-class output. A worker that found nothing relevant must be able to say so. Most failures of this shape start with a worker that had no good document and produced a summary anyway.

What would have to be true for it to self-heal

Only if some downstream component has independent access to ground truth. Nothing inside the model layer provides that, which is why the fix is a deterministic check against retrieved text rather than another agent reviewing the output.

When not to use several agents

If the four subtasks are not genuinely independent, or if their results are short enough to fit one context window, one agent with four tools beats four agents. It removes every handoff, keeps all evidence in one transcript, and is far easier to debug. The multi-agent form earns its cost only when subtasks run in parallel for long enough that wall-clock time matters, or when they need genuinely different tool permissions.