“Autonomous agent” is often used to mean “a model can do a lot of steps.” For real systems, autonomy matters when a process can act over time without causing chaos.

OpenAgents optimizes for predictable autonomy (Autonomy-as-a-Service): a contracted outcome over time, not “AI” or “tokens” alone.

• Scope: “Do X”
• Horizon: e.g. “over the next 24–48 hours”
• Constraints: budget, privacy, repo boundaries, allowed tools
• Verification: objective checks (tests pass, PR merged, receipts emitted)
• Reliability: known failure modes + escalation (“pause + ask human”)

In practice that means: scoped work (what is in and out of scope), explicit constraints (permissions, timeouts, budgets), verification-first loops (tests/builds are the judge), and legible traces so failures are diagnosable and improvements are real. The canonical output of an agent session is the Verified Patch Bundle: PR_SUMMARY.md (human-readable), RECEIPT.json (machine-readable), REPLAY.jsonl (replayable event stream). Specs: crates/dsrs/docs/ARTIFACTS.md and REPLAY.md.

Why budgets matter

Budgets are not only about cost. They are a safety boundary:

• they limit how far an agent can go if it is wrong
• they force explicit tradeoffs in routing and tool use
• they make “pay per job” markets possible

Why verification matters

In software, “truth” is downstream:

• tests
• builds
• reproducible checks

Agent narration is not a substitute.

This is why Autopilot centers a loop of: plan -> act -> verify -> iterate.