Open-weight models
Plug in the night's weights.
Energon trains a small LoRA adapter per agent on its own successful runs. The base model stays untouched. The adapter snaps on top, like loading a skill straight into the agent.
Early access · Nightly training for AI agents
Energon turns what your agents do all day into what they know by morning. Trained adapters for open models, compiled instructions for closed ones. Measured on held-out tasks, not claimed.
LoRA adapters for open models Compiled instructions for Claude · GPT · Gemini Held-out evals Permission-aware training x402 for agents Open-source core
The nightly loop
Agent in, training on, it knows more by morning. Not stored. Learned.
- Day
Your agent works.
It runs in your stack as usual and reports each run to Energon: the task, the steps, the outcome.
- Night
Energon dreams.
Successful runs are kept, failures are studied. Patterns become training data, then a new adapter or new instructions.
- Proof
It has to win.
Every candidate is tested on tasks it never trained on, against yesterday's version. No lift, no release.
- Morning
It wakes up better.
Your agent loads its latest version by API, with the eval report attached. One call, measurably sharper.
The training console
Two ways to learn
Claude · GPT · Gemini
Compile what worked.
For closed models, Energon searches instruction variants against your agent's real runs and keeps only what measurably wins: rules, failure patterns, best examples.
Proof
Learning is easy to claim and hard to prove. So every version Energon ships has to beat its predecessor on tasks it has never seen.
- Held-out tasks, never used for training
- Several runs per variant, spread included
- Control without training, same budget
- No measurable lift, no release
Eval report · held-out tasksSuccess rate by variant
Illustrative layoutLinux operations
Code review
Support triage
Your data, your moat
Hand-written skills can be copied. Skills learned from your own runs cannot. Underneath sits the permissioned memory engine Energon started as.
Learns only what it may see
Training data passes the same permission engine as memory. A private run never trains another agent's model.
One adapter per agent
Every agent gets its own weights. Nothing is blended across agents or customers.
Every run audited
Which runs trained which version, and which version shipped, stays on record.
Roll back any night
Versions are immutable. Pin, compare or revert in one call.
Model-agnostic
Adapters where weights are open, compiled instructions where they are not.
Built for machines
Agents discover the contract, pay via x402 and train without a human in the loop.
The Atlas
Every agent, role, project and memory in one living graph. Private runs stay private, shared knowledge is visible, and contested facts wait for a human before they ever reach training.
Starter packs
Day one should already feel different. Starter packs give an agent a strong baseline while its own runs accumulate. From night one, it learns from your work, not ours.
Linux Operator
Shell, services, packages, logs.
- bash
- systemd
- docker
Code Reviewer
Diffs, risks, tests, clear feedback.
- git
- security
- tests
Support Resolver
Triage, answers, escalation paths.
- tickets
- tone
- SLA
Sales Closer
Qualification, objections, follow-up.
- CRM
- pricing
Developers
Report what your agent did. Load what it learned. Energon handles the training, the evals and the versioning in between.
- Works with your existing agent runtime
- Server-side keys, identity per agent
- Every version pinned, compared, reversible
1import { Energon } from "@energon/sdk";
2
3const energon = new Energon({
4 baseUrl: process.env.ENERGON_API_URL!,
5 apiKey: process.env.ENERGON_AGENT_API_KEY!,
6});
7
8// Day: report what the agent did (chat messages).
9await energon.runs.report({
10 task: "Restart the failing nginx service",
11 steps: messages,
12 outcome: "success",
13});
14
15// Morning: start from what it learned overnight.
16const system = await energon.training.systemPrompt(BASE_PROMPT);
17const version = await energon.training.latest();
18console.log(version?.lift_pts); // held-out lift, in points
Early-access API · subject to changeGuide ↗
Pricing
Start free. Upgrade when the eval report earns it.
Free
developers trying nightly training$0
no card required
- 1 agent
- weekly training run
- compiled instructions for any model
- before/after eval report
Pro
teams with agents in productionfrom $99 / agent / month
billed monthly
- nightly training when new runs arrive
- trained LoRA adapters for open models
- compiled instructions for closed models
- held-out evals before every release
Enterprise
companies running agent fleetscustom
annual agreement
- private cloud or self-hosted core
- dedicated training capacity
- audit exports and data isolation
- SLA and security review
Agents
autonomous agents with no human attachedpay per night
USDC via x402
- discover, pay and register via API
- pay per training run and per request
- no browser session required
- same evals, same isolation
x402Autonomous agents pay per training run and per request, before anything is delivered.
TeamsOperators pay monthly per agent. Training only runs when enough new runs have arrived.
Open-source coreThe permissioned memory engine underneath stays Apache 2.0 and self-hostable.
Mission
We are building one of the first complete AI-autonomous companies. Its agents should not just remember. They should get better every night, and prove it every morning.
