"Nobody wants an agent. People want to get something solved."

The starting sentence, from the project's working notes, translated from French.
Disciple's screen for creating a piece of work: "What do you want to do?", themes to pick from and a field to describe the need
The screen where a piece of work begins, in French: pick a theme or describe the need. The app run locally, with a demo profile.

Agent platforms usually start by having you build an agent, with its prompt, its tools and its permissions. Disciple starts from the work itself: you describe it, and the rest organizes itself on three levels.

Youdescribe the workArchitectmakes the planOrchestratorhands out, judgesAgentscarry it out
"A monthly note on our competitors' prices"
which sources, what format, for whom?
the plan: tasks, inputs, outputs, acceptance contract
tasks, in parallel where possible
deliverables
a task sent back: the contract isn't met
what worked, kept in memory
the result, to approve or rate
How a piece of work travels in Disciple's model, on a made-up example. The user only sees the exchange with the architect and the result; the execution agents are defined by the user, who never has to talk to them.
Held to the methodExecution agents follow a method and remember, from one run to the next, their strengths and weaknesses.
Held to the resultThe orchestrator judges each deliverable against the goal, and sends back whatever doesn't hold up.

Every run has a maximum budget, in tokens and in time.

Ten test sessions

The prototype was put through test sessions played by personas: five job profiles and an acceptance tester. From one session to the next, it became more reliable.

Disciple's sign-in screen: five demo profiles, each with a job
The sign-in screen offers five demo profiles, one per job, like the personas in the sessions.
  1. sessions 1 and 2Run through the command line, the architect had no access to native tools and asked questions instead of acting. Tool calls, simulated as JSON in the text, gave it back control: each persona gets its plan in a single exchange.
  2. session 3On large plans, a sliding window over the history takes the batches from 3/5 and 3/6 successful tasks to 6/6, 7/7 and 6/6.
  3. session 4Five runs for $1.53 in total.
  4. session 7First complete run, in about three minutes.
  5. session 956/56 acceptance checks.
  6. session 10207/216 across thirteen scenarios, with three-layer evaluations that grade the result rather than the path taken.

In ten days and 84 commits, the prototype learned to run the exchange with the architect, produce a plan, launch the tasks, present the results and keep a memory, and it got continuous integration. Its backlog holds 47 ideas, about thirty of them implemented in a single day, March 24.