Multi-Agent Systems

1 / The Architecture Problem

Multi-agent systems was not a technology I adopted from a library — it was an architectural pattern I implemented across Conclave, Foundry, and Vigil. Each project explored a different facet: cooperative orchestration, specialized evaluation, and safety verification.

2 / Cooperative Orchestration in Conclave

In Conclave, multiple model agents participated in structured debates. The engineering challenge was context unification — ensuring that Agent B understood what Agent A had accomplished within a shared canonical state. This required translating prompt context across distinct model tokenizers and managing a unified context window across agent hops.

3 / Specialized Evaluation in Foundry

Foundry used LangGraph to orchestrate three specialized agents — each responsible for a different dimension of startup architecture evaluation. The agents operated on shared state with explicit handoff contracts, and their decision histories were checkpointed to PostgreSQL for review. The consensus mechanism was not majority-vote — it was a structured state aggregation that preserved each agent's reasoning trace.

4 / Safety Verification in Vigil

In Vigil, the multi-agent dimension was the evaluation harness rather than the application itself. Agent Evaluation harnesses tested agent behavior inside isolated Docker sandboxes, verifying that agents produced correct side effects and did not violate security boundaries. The architectural inversion — testing agents rather than orchestrating them — required fundamentally different design decisions than cooperative systems.

5 / Key Insight

The most important lesson from building three multi-agent systems: orchestration and evaluation are complementary but opposed disciplines. Orchestration assumes cooperative intent and optimizes for smooth state handoffs. Evaluation assumes unverified behavior and optimizes for containment and measurement.