All AI videos

AI video

Stop Telling Agents How — Dan Barrett, Raymond Weitekamp & Jose Montes de Oca, OpenProse

The OpenProse team shows multi-agent workflows written as plain-markdown contracts and compiled by a harness, with demos for review, event-triggered jobs, and model documentation.

AI Engineer · 2026 featured video

Watch on YouTube

What this video shows

The speakers separate what a workflow should achieve from how a specific harness runs it. A contract written in markdown declares the goal, inputs, and checks, and a compiler maps it onto a harness such as Claude Code or Codex. A pull request or webhook can trigger the same contract, which turns an ad hoc prompt into a repeatable job.

Their demos include a strict code reviewer and a bank model-documentation flow. They also benchmark runs against the contract and compare the idea with DSPy. The approach improves reuse and reviewability, but a contract still depends on the harness it compiles to, and the workshop does not publish an independent comparison of contract runners. Treat the portability claim as a design goal to test on your own harnesses.

Read the project's OpenProse documentation for the contract format shown in the demos. Compare the idea with DSPy, the programmatic prompt-optimization framework the speakers discuss.

What you will learn

  • Write the goal, inputs, and checks for a workflow before choosing the harness that runs it.
  • A contract that a webhook or pull request can trigger turns a one-off prompt into a repeatable job.
  • Evaluate every run against the contract so a passing run has evidence, not just a plausible output.
  • Declarative workflows still depend on their compiler and harness, so test portability rather than assuming it.

How to apply this safely

  1. Pick one repeated agent task and write its goal, inputs, and pass conditions as a short markdown contract.
  2. Choose a harness and compile the contract, then run it on a fixed task set.
  3. Add a trigger such as a pull request or a webhook and confirm it produces the same result as a manual run.
  4. Score each run against the contract and keep the outputs that pass as regression cases.

Important limitations

  • The workshop is presented by the project's own team and demonstrates their tool rather than an independent comparison.
  • The contract format is the project's design choice. Its portability across harnesses is demonstrated, not measured against alternatives.
  • A declarative contract can still encode a wrong task, so clear goals do not replace checking whether the goal is correct.

Sources to check

  • OpenProse The project that compiles markdown contracts into agent workflows.
  • DSPy The framework the speakers compare with contracts.

Continue learning on Learnetto

AI evals guide

Score each contract run against fixed pass conditions.