Agentic Autonomy Levels

By Addy Osmani (2026). Originally published on Substack.

A two-axis framework for AI autonomy (Agency vs. Orchestration), mapping the 6 autonomy levels and their safety mechanisms.

Core Extraction Summary

1. Named Frameworks & Patterns (Author’s Exact Words)

“Calibrated Autonomy”: Matching the autonomy level granted to an agent with the task’s blast radius, reversibility, and available verification evidence.
“Two-Axis Autonomy Model”:
- Agency axis (how far one agent goes): Low (suggest/wait) -> Mid (scoped task/evidence) -> High (autonomous goal pursuit).
- Orchestration axis (fleet coordination): Low (single thread) -> Mid (isolated worktrees) -> High (management-by-exception factory).
“Six Autonomy Levels”:
- Level 0: Assist (autocomplete, inline edits, pair chat; local verification).
- Level 1: Supervised action (agent runs commands with human approval per action).
- Level 2: Scoped task delegation (bounded task with explicit goal and constraints; test evidence).
- Level 3: Goal-driven autonomy (agent cycles plan-act-test until measurable stopping condition holds).
- Level 4: Parallel delegation (multiple agents working on isolated task slices across worktrees).
- Level 5: Managed-by-exception orchestration (manager agent dispatches, monitors, and aggregates workers; escalates only exceptions).
“Approval Fatigue”: Failure mode at Level 1 where repetitive permission prompts cause humans to auto-approve without reading.
“False Parallelism”: Failure mode at Level 4 where parallel agents edit overlapping code, causing merge collisions.
“Fleet Cosplay”: Running dozens of parallel agents while manually micromanaging dependencies instead of automating coordination.
“Summary Substitution”: Accepting an agent’s markdown summary in place of rigorous evidence review.

2. Decision Rules

When selecting an autonomy level for a task, evaluate its blast radius and reversibility: assign Low autonomy (L0-L1) to high-risk/irreversible changes, and reserve High autonomy (L3-L5) for bounded, measurable, easily reversible tasks.
When scaling agent autonomy, move up one axis at a time (increase agency on single tasks before attempting multi-agent orchestration).
When configuring Level 3 goal loops, provide automated, non-vague stopping criteria (e.g., zero linter errors, passing test suite) rather than subjective instructions.

3. Anti-Patterns & Failure Mechanisms

“Approval fatigue”: Clicking through dozens of tool execution prompts indiscriminately.
“False parallelism”: Spawning multiple agents without strict file boundaries, creating merge conflicts.
“Fleet cosplay”: Pretending to operate a factory while manually coordinating every step.
“Summary substitution”: Reviewing high-level agent text rather than inspectable evidence bundles.

4. Quantitative Claims & Qualifiers

Anthropic Claude Code research: agents ask for clarification 2x more often than users interrupt; experienced users (~750 sessions vs <50) rely more on auto-approvals and interrupt only by exception.
Steve Yegge’s single-axis ladder (“Welcome to Gas Town”) serves as single-agent baseline; two-axis model is required for multi-agent fleets.

5. What the Source Does NOT Claim

Does NOT claim that Level 5 orchestration is superior for all tasks; explicitly notes that Level 0 and Level 1 remain optimal for delicate, exploratory, or high-risk engineering work.

Full Text

In most conversations about agentic engineering, the action has changed from prompting to operating.

The core question about an action is always: what level does this task deserve, and what verification makes that level defensible?

The Two-Axis Framework: Agency vs. Orchestration

Agency Axis: How far away from yourself do you let a single agent go?
- Low: Suggests candidate actions and waits.
- Mid: Scoped task with continuous progress reports and evidence.
- High: Goal-driven experimentation with automated stopping conditions.
Orchestration Axis: How many agents run and who coordinates them?
- Low: Single thread.
- Mid: Multiple isolated worktrees.
- High: Manager agent running continuous queues (“management by exception”).

The Six Autonomy Levels

Level 0 (Assist): Suggestions, autocomplete, inline edits. Human decides everything.
Level 1 (Supervised action): Agent executes commands with human pre-approval. Watch out for approval fatigue.
Level 2 (Scoped task delegation): Bounded task with clear acceptance criteria. Verification via test evidence.
Level 3 (Goal-driven autonomy): Agent iterates until measurable condition is met (/goal, /loop).
Level 4 (Parallel delegation): Fleet of isolated agents across worktrees. Requires clean task decomposition.
Level 5 (Managed-by-exception orchestration): Event-driven factory. Manager agent delegates, verifies, retries, and escalates exceptions.

Failure Modes & Antidotes

Approval Fatigue -> Delegate boundary checks to auto-reviewers or move to scoped sandboxes.
False Parallelism -> Isolate checkouts via git worktrees and strict file ownership rules.
Summary Substitution -> Bundle complete evidence packets (diff, tests, logs, screenshots).
Fleet Cosplay -> Encode coordination rules into skills and reduce WIP limits.