# autoresearch
Autonomous experiment loop for architecture and training search.
## How it works
1. **program.md** defines the experiment scope and success criteria.
2. The agent reads in-scope files, branches off `master`, and initializes **results.tsv**.
3. Each iteration edits `train.py` (or related config), runs training under a fixed time budget, and logs `val_bpb` + memory to **results.tsv**.
4. Good runs get committed; bad runs are reverted or marked `crash` in the log.
## Quick start
```bash
uv run prepare.py
uv run train.py
```
See **program.md** for the full agent playbook and **results.tsv** for the experiment history.Multi-surface intelligence
One intelligence layer for every way work moves.
Start in the IDE, delegate to Cloud, automate through CLI, and approve from Mobile. 4RGED keeps model policy, workspace context, and review state attached to the work.
- Build4RGED IDETurn intent into code with repo context, terminals, diffs, and review in one workspace.
- Delegate4RGED CloudSpin up parallel agents for long-running work, inspect artifacts, and bring finished patches back.
- Automate4RGED CLITrigger governed agent runs from the shell, scripts, and CI while staying tied to the same runtime.
- Steer4RGED MobileApprove risky steps, monitor progress, and redirect runs when the team is away from the keyboard.
Vendors sell you more tokens. 4RGE helps you spend fewer.
4RGE is built to reduce your spend, not grow it. Break down consumption by model to find inefficiencies, by engineer to spot outliers — same outcomes, fewer tokens.
Agent effectiveness across your org
See which teams and engineers are shipping the most with agents, where adoption is accelerating, and where enablement can close the gap. Turn your power users into the playbook for everyone else.
Daily, weekly, and monthly active engineers
Sessions, messages, and active users over time — broken down by team and repo so you can tie agent activity back to real work.
File changes, commits, and shipped PRs
Connect agent usage to the work you actually merged: file changes, commits, and PRs delivered with an agent in the loop.
Vendors sell you more tokens. 4RGE helps you spend fewer.
4RGE is built to reduce your spend, not grow it. Break down consumption by model to find inefficiencies, by engineer to spot outliers — same outcomes, fewer tokens.
Agent effectiveness across your org
See which teams and engineers are shipping the most with agents, where adoption is accelerating, and where enablement can close the gap. Turn your power users into the playbook for everyone else.
Daily, weekly, and monthly active engineers
Sessions, messages, and active users over time — broken down by team and repo so you can tie agent activity back to real work.
File changes, commits, and shipped PRs
Connect agent usage to the work you actually merged: file changes, commits, and PRs delivered with an agent in the loop.