Production Agent

The problem

Open-ended agent loops wander, repeat themselves, and take actions no one signed off on. A production agent needs explicit state, bounded steps, and human approval on side-effecting actions.

System architecture

A LangGraph state machine where each node is a deliberate step: the agent reads typed state, may call a whitelisted tool, and routes to the next node. Side-effecting tools are gated behind an approval node, and a step budget bounds the loop.

Workflow

• A request enters with typed initial state.

• The agent plans the next step against the current state.

• It calls a whitelisted tool, or routes to an approval node for side-effecting actions.

• A human approves or rejects gated actions before they run.

• Tool results are written back into typed state.

• The loop continues until a stop condition or the step budget is reached.

Stack

• LangGraph for the agent state machine

• A typed state schema for each node

• A whitelisted tool registry with input validation

• An approval node for side-effecting actions

• A step budget and timeout to bound the loop

What gets logged

• The initial request and typed starting state

• Each node entered and the transition that led there

• Tool calls with validated inputs and returned outputs

• Approval prompts and the human decision

• State snapshots after each step

• The stop condition or budget that ended the run

Where evals run

Evals run against recorded agent trajectories — scoring whether the agent reached the goal, stayed within budget, and only took approved side effects — plus regression replays when tools or prompts change.

Failure modes

• The agent loops without progress — the step budget halts it instead of running indefinitely.

• A tool receives malformed input — input validation rejects it before the side effect.

• A side-effecting action runs without sign-off — the approval node blocks it by construction.

• A tool errors mid-run — the failure is written to state and the agent routes to a recovery or stop path.

What this demo proves

That an agent can be made bounded and governable — explicit state, whitelisted tools, and approval gates instead of an unbounded autonomous loop.