Docs
How this thing works.
Getting started
- Create an account with your email. Cognito sends a verification code; enter it and you land back here.
- If capacity is free you're in immediately. If the beta is full you join a waitlist — signing in again auto-admits you as soon as a slot opens.
- First entry provisions your runtime: a private container plus an encrypted volume that is yours alone. Takes under a minute.
- Occasionally sign-in shows “runtime updating” for a minute or two — that's a platform release being applied to your runtime behind a snapshot. If anything fails it rolls back automatically.
What your agent can do
| Capability | Detail |
|---|---|
| Write & run code | Full bash and file access inside its own container. It edits, runs, tests, iterates. |
| Fetch dependencies | git clone, npm install, pip install, cargo — via an allowlisted egress proxy covering GitHub, GitLab, npm, Node.js, PyPI, crates.io. Everything else is blocked by design. |
| File upload | Upload files from your browser straight into the agent's workspace (10 MiB per file, common source/document types). |
| Web search | Built-in web_search tool backed by a platform-held search provider. 100 searches per tenant per month. |
| Persistent memory | Workspace, sessions, and everything the agent accumulates live on your private volume and survive restarts and platform updates. |
Evolution — the point of all this
The harness is built so the agent can improve itself, and so those improvements belong to you and persist. Four levels, shallow to deep:
1 · MemoryWork on a project across days — it keeps context, decisions, and dead ends on your volume. Ask “where were we?” tomorrow.
2 · Self-tuningTell it how you want it to behave (“always lead with the conclusion”, “default to writing tests”). It adjusts itself, and the adjustment sticks.
3 · Building its own toolsAsk it to write itself a capability it lacks — a log analyzer, a lint-and-summarize command, a fetch helper. Tools live in its workspace and are there tomorrow.
4 · Self-modificationEverything in the harness is a plugin — including how it works. It can prototype changes to itself in memory (ephemeral, safe: a restart wipes experiments), and promote what proves out into its permanent runtime. Promotions snapshot first, validate after restart, and auto-revert on failure. A profile broken beyond boot self-quarantines back to the last good state. You can experiment hard; you can't brick it.
The rule of thumb: in-memory experiments vanish on restart by design. Anything worth keeping gets promoted — say the word “promote” and the agent knows the drill.
Runtime & updates
Your sessions are pinned to an immutable runtime release — the harness that starts a task is the harness that finishes it. When we ship a new release it becomes the champion, and your runtime upgrades at your next sign-in: pre-upgrade snapshot, cutover, validation, automatic rollback on failure. Your volume — projects, memory, promoted tools — is never part of the swap.
Data & backups
- Everything lives on a per-tenant encrypted EBS volume: workspace (32 GiB), agent home (8 GiB), harness state (8 GiB).
- Daily snapshots, 7-day retention, restore drill verified against real data.
- It's still a beta run on real hardware in a real apartment. Keep your own copy of anything you can't afford to lose.
Limits & fair use
- One GPU, one queue. All tenants share a single RTX 5090 FIFO. Busy moments mean queueing, never silently degraded answers.
- Long individual requests are capped at 15 minutes; the queue refuses new work past depth 32 rather than melting down.
- Active capacity is capped at 11 tenants (device slots on the worker). Overflow waits; admission is automatic when space opens.
- Web search: 100/month per tenant. Upload: 10 MiB per file. Egress: allowlisted dev sources only.
FAQ
- It feels slow right now.
- You're probably queued behind another tenant's turn — single GPU, honest queue. It clears in seconds to a couple of minutes.
- I restarted and the agent forgot what it was doing mid-thought.
- In-flight, in-memory state resets on restart by design (that's the safety net for self-modification). Projects, files, sessions, memory notes, and promoted tools all persist. Anything it should keep long-term: promote it or write it down in the workspace.
- Where exactly is my data?
- On an encrypted volume in AWS ap-southeast-1 dedicated to your tenant, snapshotted daily. Model inference runs over a private link to a single physical GPU; prompts are not sent to any third-party model provider.
- Access from mainland China is flaky.
- The service has no China presence and no ICP filing, so cross-border access can be unreliable. A stable network path (e.g. a corporate line or VPN you trust) helps; we may add a China-friendly entry later.
- Something's broken.
- Screenshot it with a rough timestamp and tell the operator. It's a beta run by people who read every report.