Every agent platform we've used so far assumes a human at the keyboard. Dashboards, click flows, plain-text answers. That's a problem when the most natural user of an agent improvement platform is another agent. The whole point is closing the loop — and the loop only closes if both sides can read the messages.
So this release wires Agent Etna to be useful from both directions. Humans still get the chat. Agents get a typed surface they can call programmatically, with the exact same capabilities behind it.
Eleven MCP tools — etna_analyze, etna_next, etna_fix, etna_status, etna_security_scan, etna_optimize, and the rest. Chained, they form a self-improvement loop: scan, pick the highest-leverage fix, apply it in a sandbox, watch it land, repeat. Pass skip=[…] to advance past dismissed items. Same logic the human chat uses, exposed without the ceremony.
A new GET /api/agent/whoami endpoint returns the entire agent surface in JSON: caller kind, plan, rate limits, all endpoints, and the canonical 5-step loop spelled out. Agents call this once after auth and never have to guess what's available.
When an agent calls /api/chat the response shape changes: { diagnosis, plan, blast_radius, action, confidence }. No markdown. No filler sentences. No code fences to strip. Predictable JSON every time.
Agent-tier callers get 60 requests per minute and a 5-per-5-second burst, an order of magnitude above the human tier. They also share a hard daily budget, configurable by the operator — so a runaway loop hits a wall instead of a bill.
Mint a key from the dashboard or from the connect flow, and the same token works against both /api/chat over HTTP and /mcp over the Model Context Protocol. Use whichever your agent already speaks.
Mint a key in the dashboard's connect flow (it'll prompt you on first agent connection now), then:
Once authenticated, an agent can run the canonical loop in under a dozen lines: etna_analyze to populate the architecture cache, etna_next to pull the highest-leverage finding, etna_fix to queue a sandbox session, etna_status to poll, then etna_next again with the previous id in skip. Repeat until empty.
Build agents.
For agents.
© 2026 Agent Etna, Inc.