ControlForge isn't a model wrapped in a loop. It's a deterministic execution runtime your agent calls to actually do the work — stateful, auditable, and batteries-included. Every number below is measured against the live runtime.
The non-deterministic part never executes anything directly. Five stages turn a probabilistic proposal into a bounded, recorded action.
The agent introspects a typed tool surface — every capability it can call — and proposes the steps against what actually exists. No guessing, no hallucinated APIs.
Proposed logic is validated against the runtime before it runs. Malformed or unsupported steps are rejected at the door — they never touch a real system.
Scheduling, role-based permissions, and explicit limits constrain what any step is permitted to do — and when.
A deterministic engine runs the steps — the same way every time. State persists across steps and restarts. Hot paths run on a compiled bytecode VM.
Every value and action lands in the historian and event stream, time-stamped. A deterministic state hash and full diagnostics make any run replayable and provable.
The substance under the story — what your agent can call, and what keeps it honest.
Most agent runtimes stop at the edge of the screen. ControlForge doesn't have to — when a task needs to touch a device, a sensor, or a machine, it already speaks the protocols. Not the pitch; just a door the others don't have.
ControlForge is in private access. Tell us what you're building.
hello@controlforge.tools