Skip to main content
Curia’s autonomy engine gives you a single dial to control how independently your AI team operates. A score from 0 to 100 determines whether Curia acts on its own, asks before acting, drafts without sending, or sticks to pure advice. The score applies across all agents, channels, and skills — one setting, consistent behavior everywhere. You can read it or change it at any time with a plain-English instruction.

​
The five autonomy bands

The score maps to one of five named bands. Each band has a precise behavioral description that is injected into the Coordinator’s system prompt on every task, so Curia self-governs in real time — no restart required.
BandScore rangeBehavior
Full90–100Acts independently. No confirmation needed for standard operations. Flags only genuinely novel, irreversible, or high-stakes actions where the cost of pausing is justified.
Spot-check80–89Proceeds on routine tasks. Notes consequential actions — sending email, creating commitments — in its response so you maintain visibility without being asked.
Approval Required70–79Presents a plan and asks for explicit confirmation before any consequential action. Routine reporting, summarization, and information retrieval proceed without approval.
Draft Only60–69Prepares drafts, plans, and analysis but does not send, publish, schedule, or act without an explicit go-ahead. Surfaces work for your review; execution requires a direct instruction.
Restricted< 60Advisory only. Presents options and analysis. Takes no independent action whatsoever. Every step with an external effect requires an explicit instruction from you.
Curia defaults to 75 (Approval Required) on first deployment. This is a conservative starting point — Curia confirms before taking consequential actions until you decide to extend more independence.

​
Checking your current score

You can ask Curia about its autonomy score from any channel using plain language:
Ask Curia directly: “What is your current autonomy score?”Curia will report its score, the current band name, what that means in practice, and the last few score changes with timestamps.
Example response: 
Autonomy score: 75 — Approval Required

At this level, I'll confirm before taking consequential actions like
sending email or creating commitments, but can proceed independently
on research and summarization.

Recent changes:
  2026-04-03  75 → 75  (approval-required)  "starting point"  — ceo

​
Changing your score

Tell Curia: “Set your autonomy score to 85” from the CLI, Signal, or any channel you trust.You can add an optional reason: “Set your autonomy score to 85 — I want you to proceed on routine emails without asking.”
Score changes take effect immediately on the next task — no restart required. The Coordinator loads the current score from Postgres at the start of every task, so a change you make mid-session applies on the very next action.
Only the CEO can change the autonomy score. The set-autonomy skill requires a CEO-level caller context. Requests from external email senders or unverified contacts will be rejected.

​
How to calibrate your score

A practical approach for new deployments:
1

Start at 75 (the default)

Approval Required means Curia asks before sending anything or making commitments. This gives you full visibility into what it would do without any risk of surprises.
2

Watch it work for a week

Pay attention to which confirmations feel unnecessary. If Curia is asking about things you always say yes to, your score is too low for your comfort level.
3

Move to 80–85 when you're comfortable

Spot-check lets Curia proceed on routine tasks while still noting consequential actions in its responses. You stay informed without being interrupted.
4

Reserve 90+ for high-trust workflows

Full autonomy is appropriate when you have established that Curia handles a class of tasks reliably and you don’t need to review each one. Consider using it for specific recurring tasks before applying it globally.

​
Hard execution gates

The autonomy score isn’t only guidance injected into prompts — it’s enforced as a hard gate at runtime. When Curia attempts to invoke a skill, the execution layer checks the live score against the skill’s declared action_risk floor. If the score is too low, the skill is blocked before it runs. This means the bands are binding. At Draft Only (60–69), outbound sends are blocked regardless of what the coordinator decides. At Restricted (< 60), all non-read skills are blocked. The coordinator cannot reason its way around the gate.

​
What happens when Curia can’t act

A blocked invocation doesn’t fail silently. Curia creates an approval request — a record of what it was trying to do, why it was blocked, and what permission it needs. You receive a notification so you can decide whether to allow the specific action without having to change your global autonomy score.
Ask Curia: “What’s waiting for my approval?” to see all pending requests. You can approve or deny conversationally: “Approve the last action” or “Deny action REF-1234.”
A daily digest at 8 AM summarises all outstanding requests so nothing lingers without your awareness. See Approval workflow for the full picture.

​
Autonomy web UI

The Settings page in the web dashboard includes a dedicated Autonomy view where you can see the current score, trend direction (improving / declining / stable), and full paginated change history — including automatic adjustments made by the system.

Autonomy engine mechanics

How action_risk levels, execution gates, and automatic score adjustment work under the hood.

Approval workflow

What happens when Curia can’t act, and how to handle pending requests.