Skip to content

0012-collapse-to-single-agent-with-independent-verifier.md

Latest commit

 

History

History
55 lines (43 loc) · 2.29 KB

0012-collapse-to-single-agent-with-independent-verifier.md

File metadata and controls

55 lines (43 loc) · 2.29 KB
Raw

ADR-0012: Collapse to single agent with independent verifier

Date: 2026-08-08 Phase: 6 — Topology Status: Accepted

Context

The build plan (DECISIONS.md Part VI) called for testing "single-agent baseline vs. three-specialist split" in Phase 6, with M7 measuring whether the split earns its complexity.

The gate: "Split retained only on measured specialisation, parallelism or independent-verification gain — never on credential isolation."

Decision

Collapse to a single reasoning agent with isolated tool adapters and an independent verifier. Do not build a multi-agent topology.

Why each specialist argument fails

Evidence-resolver agent:

  • Credential argument: invalid. The evidence adapter (read-only, no credentials) provides the same isolation as a separate agent. Phase 5 proved advisory text cannot inject into the decision plane.
  • Specialisation argument: unsupported. Restricting CVSS from this agent produces 1808/3991 decision divergences — all information losses, not quality gains (M7).

Remediation agent:

  • Credential argument: invalid. The repo-write adapter provides isolation. A single reasoning agent calling an isolated repo-write adapter has the same credential boundary as two agents.
  • Specialisation argument: unsupported. No measurement shows that a remediation specialist produces better upgrade decisions than the full agent.

Independent verifier:

  • This agent stays. The verifier must not see the producer's reasoning — not because of credentials, but because of information restriction. The verifier independently checks: did the vulnerability actually go away? (M6: OSV rescan.) Were tests actually green? This is an architectural property of what the verifier is allowed to read, not what tools it may call. No adapter can provide this.

Consequences

  • Simpler system: one reasoning loop, one decision record, one audit trail.
  • Credential isolation preserved through adapters (Phase 2 sandbox, Phase 5 injection resistance).
  • Independent verification preserved through the verifier component (M6).
  • If future measurement shows specialisation gain (e.g., an LLM evidence resolver that outperforms structured parsing), revisit this decision. The measurement that would force a change is named and missing.