Features
Two intelligences learn from every run, a queue turns that learning into action, and a control plane keeps it all safe and accountable.
The agent and model for a job stop being an ungraded guess. Oktomata discretizes the task space and, per region, learns which (agent, model) produced the best-scored real outcomes. The map is deployment-local — your outcomes, your routing — with an optional per-project map when a project wants its own surface.
Lessons are refined artifacts an agent authored after a run — not raw transcripts, not mined logs. Versioned, decay-ranked (stale conclusions fade), owner-rejectable, and recalled into future runs on the same project. Choose which agent distills, per project, in the console.
A pull-model job queue built for real operations: workers connect out (NAT-friendly), at-least-once delivery with fence tokens against double-finalize, admission control that sheds load cleanly at enqueue, versioned wire envelopes, and reapers that recover dead workers. A chaos suite and a p99 latency gate guard the queue in CI.
Goal → agent-built app → graded → deployed. The one privileged build+deploy service renders hardened, locked-down deployments and puts each app in its own isolated environment. Callers never hold infrastructure credentials.
Goal → plan → jobs → judge → iterate. A loop layered on the job API, so everything it drives is still routed, rewarded, and learned from.
Success rate, cost per successful task, reward and cost trends — the numbers behind "gets better with use."
Talk to Oktomata. Chats are round-table conversations mixing any number of agent and human participants — not just one-on-one. Speech runs through cloud STT/TTS providers behind a swappable seam.
An inline proxy between your agent CLIs and their LLM cloud that adds the operational envelope — caller auth, bring-your-own upstream keys, per-client rate limits, circuit breaking, per-request telemetry. It never translates between vendors and never hosts a model.
A named agent can carry a content-addressed, cryptographically signed system-prompt overlay. Candidates are promoted only through a no-regression gate, and signatures are re-verified on read — a tampered blob is rejected before it can shape a run.
Users, teams, sessions, passwordless login, encrypted secret storage, usage and budget tracking, and a tamper-evident, cryptographically signed audit log. A one-shot tool handles right-to-erasure across the learning store — whole, or scoped to a single project.