Responsible Autonomy

Preview

Where autonomy is appropriate, and where it is not.

Autonomy should be proportional to reversibility. The question is not "can the agent do this?" but "what happens if it does this wrong at 3am with nobody watching?"

Autonomy ladder#

Level	Agent authority	Appropriate when
Observe	Read and report only	Building trust in a new workflow
Draft	Produce work for a human to submit	Communication and external-facing output
Act reversibly	Execute changes that can be undone	Repository work, analysis, internal tooling
Act with gates	Execute, pausing at declared effects	Operational workflows with external writes
Act autonomously	Execute within hard budgets	Well-characterized, bounded, monitored tasks

Decisions with legal, medical, employment or safety consequences for a person.
Irreversible financial actions above a threshold a human would notice on a statement.
External communication attributed to a named individual.
Anything where the failure mode is invisible until it compounds.
Workflows where nobody has agreed to be accountable for the outcome.

Start at Observe. Move up a rung when the trace record shows the agent has been right for reasons you can inspect — not merely right.

Last updated 2026-09-16

Autonomy ladder#

Where we recommend against autonomy#