Write the goal. Approve the plan. Decide at the checkpoint.
The agent turns a goal written in plain words into a plan of steps. Nothing runs until a person has read the plan and approved it.
A task on the demo set: the goal, the plan and the stored trace of each step.
A plan you can read
Every step is one of eight tools: review, read the answers, compare with the playbook, checkpoint, ask the person, research, redlines, memo. The agent cannot invent a ninth.
It stops for a person
At the checkpoint the agent shows what it found and waits. A question it cannot settle goes to the person, and the task does not continue until it is answered.
A trace that is kept
Each step is stored with what went in and what came out. The task ends in a draft memo that counts how many of its statements rest on a verified quote.
Keep a plan that worked
A plan can be saved under a name as a procedure and run again on another set of contracts, with the same steps and the same checkpoint.
- Code adds every step the goal’s words call for that the model left out.
- One task runs at a time.
- A restart of the app marks a running task as failed, and says so.
The agent proposes. A person decides.
It does not send, sign or file anything. Its last step is a draft for a lawyer to read.