- Author
- Jina Yoon, PostHog newsletter
- Published
- Jul 27, 2026
- Kind
- Company newsletter post proposing a decision framework
- Original
- newsletter.posthog.com/p/agent-autonomy
- Subject
- Harnesses and agent loops
- Themes
- Handing over control
What it is
Jina Yoon writes the PostHog newsletter. The post gives a simple guide for choosing how much autonomy an agent should have.
What it says
- Trust in agents should depend on the task, not on how good the model is.
- Two questions set the ceiling. Is the work easy to check, and cheap to undo?
- The post defines four levels, from agent as assistant up to self-driving mode.
- Examples include a feature flag engine migration at Level 0. A Rust SQL parser rewrite sits at Level 2.
- To raise autonomy, it suggests smaller tasks, LLM judges and guardrails in code.
- Self-driving tasks are still few, such as dependency bumps, lint fixes and added test coverage.
What evidence it gives
The evidence is a handful of PostHog anecdotes, tied to Levels 0, 1 and 2. The post gives no data, benchmarks or failure rates. The framework itself is offered as an argument.
Interests and limits
PostHog sells software and the post promotes its agents, models and job openings. The framework was not tested against other cases. A footnote says scale is not a factor in the levels. The levels cover mainly code tasks.
Connections
These links are this wiki's own reading, not claims the sources make.
- A spec the agent can't run isn't a spec yet. Yoon sets an agent's ceiling with two questions. Is the work easy to check, and cheap to undo? A spec the agent can run changes the first answer. So Lewis's check is also a way of buying autonomy.