Execution Model

The execution model is intentionally narrow:

1. the operator receives a run request
2. the capability selects the task domain
3. the engine executes with the runtime
4. verification accepts or rejects the result
5. the trace records what happened

Why this exists

This keeps the abstraction honest. The operator API talks about the run contract, not about browser clicks, prompts, or hidden planning loops.

How it works

At a high level, the runtime behaves like this:

operator.execute(...)
-> engine runs
-> candidate result is produced
-> capability default verification runs
-> per-run verify runs
-> result is accepted or retried
-> trace is returned

The current implementation uses a retry loop with a configurable maximum attempt count.

Example

Current public usage looks like this:

const execution = await operator.execute({
  goal: "Open Hacker News and extract the current top story",
  capability: "browser.extract_hacker_news_top_story",
  context: {
    url: "https://news.ycombinator.com/",
  },
  runtime: {
    page,
  },
  verify(result) {
    return result.title.length > 5;
  },
});

Notes

• The engine does the work.
• The runtime makes the work possible.
• The capability tells the engine what kind of task is happening.
• Verification decides whether the candidate result is acceptable.