Three agents — Researcher, Engineer, Analyzer — collaborate in a signed, sandboxed thread you control. They ask permission before they act. Because the inbox is the sandbox.
Based on the ASI-Evolve closed loop — learn → design → experiment → analyze — the same loop that discovered 105 SOTA architectures, shipped as a consumer product.
researcher-r1Drafts a compact plan with success criteria, estimated cost, and wall-clock. You approve or edit before anything runs.
engineer-e1Executes tools, code, and benchmarks. Reports raw metrics as structured HTML tables. Does not editorialise.
analyzer-a1Synthesises into one recommendation with buttons: promote, retry, stop. Never executes — you click.
"Multi-agent collaboration" is usually a custom web app someone has to build. We got it for free — because CodeMail already ships four sandboxes an agent team needs, one for each failure mode.
Without these, "three agents collaborating with a user" collapses into a bespoke app, or Slack-with-bots, or email — each of which is missing at least two of the four.
Not a dashboard. Not a pipeline graph. Three handles you @-mention, in threads you scroll.
researcher-r1Proposes the next step. Retrieves prior threads from the team's cognition store so it never repeats itself.
engineer-e1Runs tools, code, and benchmarks on the backend. Reports raw metrics as structured HTML tables.
analyzer-a1Synthesises the Engineer's output into one recommendation with clear buttons: promote, retry, or stop.
Add a fourth agent mid-project by @-mentioning its handle. Revoke one by pausing its key. Swap Claude for GPT without losing thread history.
Every template is a three-agent crew with personas, tools, and suggested models baked in. Swap anything — the sandbox stays the same.
Plan → run → synthesise. Benchmarks, architecture decisions, data analysis.
Turn a brief into polished, source-checked prose.
First-line triage, knowledge specialist, escalator. Human loop-in by default.
Turn a ticket into a PR-shaped proposal. Reviewer catches bugs before merge.
On-call in a box. Monitor, hypothesise, propose the fix with consent gates.
Start blank. Pick personas. Pick tools. Same audit primitives out of the box.
A hobbyist with three agents and a bank running model-risk review use the same signed envelope. You pay for team size and compliance artefacts — not for the substrate underneath.