Stop One Agent's Mistakes From Poisoning Another

Why this matters

When agents share state, one agent's mistakes, hallucinations, or half-formed reasoning can contaminate another's context — the second agent treats the first's speculation as fact. This compounds: errors propagate and amplify across the pipeline. Preventing contamination means marking the provenance and confidence of shared context (this is a verified fact vs. this is agent-2's hypothesis) so downstream agents weight it appropriately, and isolating speculative reasoning so it doesn't leak as ground truth. It's the multi-agent version of trust boundaries.

Demo

The demo tags every shared item with provenance and status (verified/hypothesis/raw) so a downstream agent can be instructed to act only on verified items and treat hypotheses as tentative — stopping error propagation.

Try it yourself

1. Run it — the hypothesis is clearly marked so a downstream agent won't treat speculation as fact.
2. Remove the status tags and instruct the decider normally; observe how it may adopt the hypothesis as ground truth.
3. Add a 'confidence' score and have the decider weight items by it.
4. Trace a deliberately-wrong fact from one agent and confirm provenance lets you find which agent introduced it.

References

• Anthropic — Building effective agents
• Anthropic — Reduce hallucinations

When the model call fails. Read the error and decide: fix the request, retry, or fall back. 400/422 (bad params, context-length exceeded), 401/403 (auth / no access to that model), 404 (wrong model id) are fatal — fix and don't retry. 429, 500/502/503, Anthropic 529 (overloaded), and timeouts are transient — retry with backoff. Watch for non-HTTP failures too: finish_reason: "length" truncation (raise max_tokens or continue), safety refusals, malformed JSON / failed tool-call parsing (validate against a schema and repair-retry), and mid-stream disconnects. Always log the provider request id with the error so you can trace it later.

Chat about this lesson