Note: Agent Board is a collateral line of QMetrika Labs, not the core. The core is biocomputation (ΔG engine, EnergyFingerprint). Agent Board was born as the infrastructure that governs those scientific pipelines, and here we tell it as a reusable exploration piece.
The problem
When the cost of producing a text, an analysis or a decision with AI falls to near zero, four problems appear that no LLM API solves on its own: ungoverned spend (agents burn tokens with no limit or per-task cost visibility), no audit trail (no verifiable record of who approved what, when and over which payload), vendor lock-in, and data leakage (sensitive data leaves toward a cloud API).
In a regulated domain —patient genomic data, with traceability and sovereignty requirements— those four problems stop being nuisances and become blockers.
The idea: the human decides, the system proposes
Agent Board is a five-column Kanban (Queued → Working → Needs Input → Review → Done) where each agent does its part and the pipeline stops at a gate until a human authorizes. It is not decoration: the broker is fail-closed — if nobody approves, nothing advances.
A token spent and forgotten is expense. A token spent, governed and captured, is infrastructure.
The pieces
- Fail-closed approval broker: the flow stops at each gate; the human decision is cryptographically bound to the payload.
- Hash-chain audit (tamper-evident): who, when, which payload, which hash. Verifiable with one command — the basis of a regulatory dossier.
- Declarative RBAC (policy.json) + an MCP Gate with a @gated decorator.
- Universal multi-LLM adapter (~60 lines): each agent uses the optimal model —frontier, specialized or local— without vendor lock-in.
- Dual cost panel (multi-LLM vs all-Opus): makes the savings per task and per model visible.
Why it was born here
It was not designed in the abstract. Agent Board is the layer that orchestrates QMetrika's variant-classification pipelines: a Genomic Auditor (cheap model) downloads and filters, a local Biophysical Engine computes ΔG profiles without data leaving the server, a Predictor scores, and a gate halts everything until the geneticist reviews. That is the most demanding use case there is —patient data, compliance, human-in-the-loop— and it is the one that validated the approach.
Where it fits
As a product, Agent Board is the governance layer for teams deploying agents over sensitive data or in regulated domains (health, legal, finance). Today it has the product built —broker, MCP Gate, RBAC, audit, 12 ADRs, a threat model with 15 threats— and a solid internal use case, but no external customers yet: it is an exploration line, not a market-validated one.