Day 13 – Multi-agent chaos (No.13)

[onestardao]

Symptom
multiple agents hit the same PDF or index at once. instead of collaboration, their contexts pollute each other. facts blur, citations drift, and answers mutate depending on which agent touched the store first.

---

ProblemMap

• No.13 Multi-agent chaos
  parallel agents run on shared intake / retrieval without isolation. their embeddings and traces collide, leading to cross-contamination instead of independent reasoning.

---

What this usually looks like

• agent A retrieves one section, agent B overwrites or extends it with different casing or schema
• citations point to spans that only one agent saw, the other invents filler to bridge gaps
• embeddings drift because each agent tokenizes slightly differently, leaving “half-views” of the corpus
• logs show answers that change depending on the query order or which agent ran first
• bug reports from users: “same PDF, but each run gives different story”

---

60-sec quick test

1. isolation check
   run two agents on the same PDF. if their traces merge or overwrite instead of staying separate, chaos confirmed.

2. index collision
   let agents ingest the same corpus independently. if embedding counts differ or retrieval jumps, isolation missing.

3. cross-contamination probe
   ask one agent for a fact, then the second for a different fact. if second agent’s answer contains the first agent’s context, your pipeline leaked.

---

Diagnose checklist

• ingestion logs show interleaved agent writes
• embeddings or retrieval counts mismatch expected corpus size
• answers mutate across runs with identical queries
• hallucinations correlate with agent concurrency, not corpus content
• trace logs lack per-agent isolation

---

Minimal fix

Goal enforce single-source trace and index isolation.

1. separate traces
   every agent run gets its own trace log and memory slot.
2. isolate index
   agents either read-only share, or build local embedding caches before merge.
3. lock ingestion
   prevent simultaneous writes on the same PDF or corpus.
4. explicit agent IDs
   tag every span / chunk with the agent identity.

---

Hard fixes when minimal is not enough

• build multi-tenant vectorstore partitions (per agent or per task)
• introduce ingestion validators that reject mixed-agent writes
• add evaluation gates (coverage ≥ 0.7 per agent before merge allowed)
• orchestrate with a coordination layer that serializes or batches agent requests

---

WFGY guardrails you can enable

• Trace isolation. each agent run gets its own WFGY semantic tree
• Index fences. enforce per-agent embedding contract before merge
• Retrieval playbook. consistency checks across paraphrases before agents share results
• Audit log. verify intake → embedding → retrieval steps per agent

in forums, label this class as No.13. the failure wasn’t retrieval itself, but multiple agents colliding on the same source.

---

Tiny script you can paste

agent isolation sanity check (toy)

class Agent:
    def __init__(self, name):
        self.name = name
        self.trace = []

    def ingest(self, doc):
        self.trace.append(f"{self.name} saw {doc}")

A = Agent("A")
B = Agent("B")

A.ingest("PDF1")
B.ingest("PDF1")

print(A.trace)  # ['A saw PDF1']
print(B.trace)  # ['B saw PDF1']
# separate traces, no cross-contamination

---

Acceptance checks before you call it fixed

• each agent’s trace log reproducible and independent
• retrieval coverage stable across agents, no cross-drift
• no hallucinations tied to query order or agent concurrency
• merge only after validators confirm per-agent integrity

---

TL;DR

multi-agent chaos happens when agents write into the same intake or index without isolation. always enforce per-agent trace and index fences. otherwise their answers pollute each other and mutate unpredictably. call it No.13.

[ProblemMap · Article Index]