Why Anthropic’s 'Agent Teams' is missing the most important part: Governance.

Source: Dev.to

The Challenge of Trust in Multi‑Agent Systems

Most multi‑agent systems rely on a shared memory bank. If Agent A hallucinates an API call and Agent B reads that memory, the mistake becomes “institutional knowledge.” Without proper checks, a single error can propagate throughout the entire team.

If you give an agent the same context for coding and reviewing, it will inherit its own biases. In such a scenario, you lose traceability:

• Which agent made a decision?
• When did it happen?
• Did it pass the required quality checks?

Without this traceability, you don’t have a team—you have a compliance nightmare.

Rigovo’s Approach to Governance

After building Rigovo Teams—37 K lines of code, eight specialized agents, and one Master Agent—we realized that the hard problem isn’t orchestration; it’s trust.

Flat Semantic Memory with pgvector

We use a flat semantic memory store powered by pgvector. This enables fast similarity search while keeping the memory structure simple.

Deterministic Quality Gates

Every memory entry must pass through 24+ deterministic gates before it can influence the team. These gates check for:

• Phantom API calls
• Secret leakage
• Structural violations

Only after clearing these checks does the entry become part of the shared knowledge base.

Our open‑source governance layer, Rigour, performs structural analysis rather than relying solely on LLM‑based heuristics. If you can’t trace a decision back to the exact agent, timestamp, and gate result, the system fails to meet production‑grade standards.

Open‑Source Governance Layer – Rigour

Rigour is released under the MIT license and has already been forked by teams such as Alibaba’s iFlow, demonstrating that production‑grade AI needs a robust governance layer.

• Repository:
• Governance service:

We’re excited to see the big players enter the space, but our focus remains on building tools for engineers who need to ship code that works in production—not just in a demo video.