Every day your team corrects it by hand. Polygramme trains a model on those corrections, tests it against the one you run today, and you decide what ships. We do the machine learning. You keep the weights.
Take one mistake. A customer asks to change a billing address and the agent sends it to general support. Someone reroutes it by hand. Today that correction is thrown away, and tomorrow the agent does it again.
Polygramme keeps it. The past requests, what the agent did, what your people fixed: that becomes a candidate model for the one job. The candidate goes head to head with the model you run today, on requests neither has seen.
Better? Deploy it. Not better? Keep what you have. The traces are yours. The weights are yours. The decision is yours.
Recursive self-improvement: an agent runs, its attempts and corrections become training signal, a new model trains on that signal, the new model runs. Repeat.
Unsupervised, that is an agent grading its own homework. Ours has a person at the gate. Test set and pass mark are fixed before training starts. A candidate has to beat the incumbent on tasks it has never seen. Nothing ships without your approval, and every cycle leaves a receipt.
The first time our loop ran unattended it rejected its own candidate. The gain was inside the noise. That is the gate working.
Conversations, tool calls, outcomes, and the fixes people make. Your agent already produces all of it. We keep it, with your permission.
A candidate open model, or an adapter on one, learns from that record. RL where the outcome can be checked, supervised training where a person showed the right answer.
Candidate versus incumbent on held-out cases, same conditions, with a confidence interval. A comparison, not a demo.
You read the comparison and approve or reject. Rejected changes nothing. Approved goes live behind the same endpoint. Your agent's code does not change.
Nobody buys reinforcement learning. They buy a receptionist that stops quoting prices, or a coding agent that stops breaking the same tests. So the loop is built once and packaged per use case: environment, scorer, held-out set, gate. Point it at your agent. You do not need an ML engineer on staff. That is us.
The language model inside a phone agent improves from its own calls. Simulated callers phone it, a judge scores every call, the loop trains on the captured turns and gates the result. Speech recognition and synthesis stay as they are. First full cycle on a dental receptionist, September 2026.
→ For voice agent teamsAn agent that migrates or repairs code inside a sandbox, scored by tests that must pass. The first unattended cycle ran on a real migration task and the gate held the line. A second recipe targets SWE-style bug fixing.
The narrow decisions inside a bigger system: which tool to call, which team gets the ticket, which sources to search. We trained a search router from its own traces and wrote up every step. Support routing follows the same shape.
Different shape of agent? Same loop. The recipe is what we write for you in the pilot.
You have an agent doing real work. You do not need a team that knows how to run reinforcement learning. Bring the task and the corrections. We bring the training.
Support agents, coding tools, document systems, vertical products. Your customers report the same failure and you would rather fix the model than rewrite the prompt again.
Internal workflows where review is expensive, mistakes repeat, data cannot leave the building, or a large API model is doing a small job at a large price.
You customize the same workflow for every client. Make that a repeatable training step instead of weeks of prompt surgery, and hand the client weights they own.
No platform subscription, no training API to learn. One fixed-scope pilot on a workflow you already run, with a metric agreed before training starts. We run the loop. It ends with a recommendation: deploy, iterate, or stop. Sometimes that is keep what you have.
About three to four weeks once usable data and access are ready. Scope and price are set per workflow, by the founders, on a call.
Talk to the foundersEverything above runs on open tools built in public. If you would rather read code than copy, start here.
rlcli is one open-source CLI (Apache-2.0) that turns any GPU into a Tinker-API training server on SkyRL and vLLM, imports OpenAI- or Anthropic-format traces, and trains with supervised fine-tuning or RL. GSPO runs as one fused forward-backward call: 23% faster steps than the two-pass pattern hosted APIs need, measured on one A100.
A Harbor task is a Docker container, an instruction and a test script. The agent works the task over many turns and the test verdict is the reward. Runs on your own Docker daemon, no cloud sandbox account.
polyloop-rl is the controller behind RSI. Pin the task pool and the held-out split, keep the tasks the model half-solves, train (RL, then hinted on-policy self-distillation on captured traces), evaluate paired against the incumbent, gate on thresholds declared up front, promote only on approval. Every cycle is an append-only directory with a receipt. Recipes are packages on top: polyvoice plugs a voice-simulation service in as environment and verifier. v0; first unattended cycles 16 and 18 September 2026, both candidates rejected by the gate.
The same server on GPUs we operate. LoRA adapters on a shared base, an OpenAI-compatible endpoint, a relay that records your agent's traces, and weights you can download. A console for runs, traces and checkpoints.