How Much to Delegate to AI Agents? PostHog's 4-Level Autonomy Guide
rseroter · x · 2026-07-29
As AI agents take on more unsupervised work, PostHog shared a practical guide for granting agent autonomy. The post argues that the key to delegation isn't how smart the model is, but the nature of the task itself.
The author proposes two core dimensions for evaluating tasks:
- Is it easy to check the agent's work?: For example, code can be easily verified via deterministic unit tests, while subjective tasks (like renaming a parameter for clarity) are harder to evaluate.
- Is it cheap to undo an agent's mistake?: If errors can be easily rolled back (like via version control), the risk is low; otherwise, it's extremely high.
Based on these dimensions, tasks are categorized into four delegation levels:
- Level 0 (Assistant): Hard to check, hard to undo. Necessary for sensitive code.
- Level 1 (Human-in-the-loop): Hard to check, easy to undo. Requires human subjective evaluation.
- Level 2 (Agent delegation): Easy to check, hard to undo. The default safe ceiling for most dev work today.
- Level 3 (Self-driving): Easy to check, easy to undo. Agents can run completely autonomously.
More from coding & agent
- Reasoning effort is a separate control from model size, not a speed throttle — alex_verem · 2026-07-29
- Rust-based jcode claims 245× faster startup and 14× lower RAM than Claude Code — Shruti_0810 · 2026-07-29
- Weaviate’s Query Agent adds GPT-5.6 Luna and Terra with 5–10% better recall — eshorten300 · 2026-07-29
- User claims Claude-built wallet monitor turned $2K into $12K in one night — Aiden_Tech_Ai · 2026-07-29
- Git hides how agent-generated code was produced, not just the final diff — haltakov · 2026-07-29
- LangChain Launches Course on Autonomous Agent Improvement with LangSmith Engine — LangChain · 2026-07-29