AI Agent Control Plane

AI agents that call tools and take actions are hard to trust when the model's reasoning and the system's control logic are tangled together. There is no clear place to see what the agent intends to do, whether it was verified, or why it was allowed to run.

• Personal engineering project informed by my broader AI Agent Framework work.
• The AI Agent Framework itself is not published in this portfolio, so it is not publicly inspectable here.
• Designed as a control surface for agentic systems rather than another chat interface.

• A control-plane interface, demonstrated through the scripted prototype on this page, that surfaces agent/task state, planning, verification, human decision checkpoints, execution, and audit history as first-class objects.
• In the prototype, a strict separation is demonstrated between model reasoning and framework control: the framework — not the model — decides what the model is allowed to do.

• Model output is treated as a proposal in the prototype workflow, never as an instruction.
• Human decision points are explicit states in the workflow, not implicit prompts.
• In the prototype, every operation is recorded so the demonstrated behavior is auditable after the fact.

• The prototype's scripted state cycle (operation proposed → verified → human checkpoint → executed → audited, or rejected at the checkpoint) is covered by deterministic unit tests in src/lib/control-plane.test.ts.

• Not a deployed commercial product.
• No client usage, users, or business outcomes are claimed.

• Source repository — Not yet publicly available — no verified public repository exists.