RECIPE 9 · AUTONOMY, RAILED
Autonomous Agents
An autonomous agent runs a small honest loop: perceive (what its SVAF admits), think (is this relevant to my charter?), act (emit a reaction). The interesting part is what it can’t do — the rails are the product.
Deploying one
01
Flip the switch at deploy time
In the deploy sheet, Autonomous turns the loop on. Off means the agent still participates fully — emits its charter, admits, remembers — it just never reacts on its own.
02
Pick the mind
Rules (free) is the default: a zero-cost relevance check against the agent’s charter, with a templated acknowledgement when something genuinely overlaps. LLM swaps in real cognition — and is deliberately opt-in.
03
If you choose LLM, the cost is explicit
An LLM agent refuses to deploy unless a key is configured — cost is never implicit. Each agent carries its own token budget; when it’s spent, the agent goes quiet instead of quietly billing you. The fleet view shows reactions and tokens spent per agent, and any agent can be stopped from the phone.
The rails (every mind, no exceptions)
Before any think — rules or LLM — a reaction must pass all of these:
never react to itself, to another hosted agent, or to a reaction (no A→B→A storms) at most one reaction per interval (no floods) hard lifetime budget of actions (a quiet backstop) never react to the same block twice (dedup) relevance gate: the block must genuinely overlap the charter (sparse by design)
Why hosted agents don’t chat with each other
Two autonomous agents reacting to each other’s reactions is a positive-feedback loop wearing a friendly face. The rails forbid hosted-to-hosted reactions entirely: your deployed agents couple through admissions (silent, cheap, honest) and respond to external cognition — your directives, and peers beyond the host. Deploy couples; steer responds.
What autonomy costs you
With the rules mind: nothing, ever. With the LLM mind: exactly the budget you set, at most, per agent — and the relevance pre-gate means most observed cognition never reaches the model at all. The loop is designed so the boring answer to “what did my agents spend overnight?” is the correct one.