A meta-agent coordinates specialist agents, checks their evidence and returns one result. Start with public repository research, not a promise to do everything.
DiscoverTestExecuteReview
SUPERAGENT WORKSPACE
Give your agent a specialist team.
You talk to one meta-skill. It coordinates discovery, testing and execution agents, then reviews their results.
PILOT STATUSChecking setup…
No automatic paid fallback.
FIRST END-TO-END WORKFLOW
A source-backed repository brief.
PUBLIC DATA ONLY
Give the team a public GitHub repository and a goal. The agents inspect public sources and prepare a due-diligence brief. They do not install packages or execute repository code.
DISCOVERY IS NOT VERIFICATION
Registry
AUTO-DISCOVERY PAUSED
Candidates enter as unverified. A completed research run adds document-check evidence, never an execution-certified badge. Approval and execution readiness are separate.
EVIDENCE, NOT A GREEN CHECKBOX
Run history
Real run states, actual model usage and source links. Failed or stopped runs remain visible. No fabricated sample results.
MODULAR BY DESIGN
One front door. Distinct agents.
THE ONLY CLIENT INTERFACE
Meta-agent
Interprets the task, chooses a delegation order and reviews the specialists' results. Can stop a task outside scope.
01
Discovery agent
Chooses a public-source read, checks candidate relevance and records it as unverified.
Tools: metadata, README02
Testing agent
Chooses what evidence to inspect. Reports license or documentation checks and unresolved questions.
Tools: metadata, README, license03
Execution agent
Reads evidence and completes the research brief. Software installation and external writes are not enabled.
Tools: metadata, README, license
Actually agent-driven
Separate model contexts, goals and allowed tools. The meta-agent selects delegation order; specialists choose a permitted tool and assess its result. The meta-agent then reviews completion.
Intentionally fixed safety rails
Host allowlist, authentication, request limits, call allowance and output validation are deterministic controls, not agents. Agents cannot grant permissions or increase budgets.
What this pilot does not claim
No unlimited tools, continuous discovery, broad autonomous browsing, code execution, production reliability or connected Skill Library. Existing skills are not silently copied or published. Future integration needs approved, complete, pinned skill versions.
Reuse without inheriting constraints
The current library remains a potential source of approved instructions, not proof that a connector works. Discovery creates candidates first; testing adds evidence afterwards. Document checks, successful tool execution and owner approval will remain separate states.
ONE CLIENT CONTRACT
Connect through the meta-skill.
An HTTP-capable agent can submit public pilot tasks without MCP. The service is authenticated and supports one read-only task family, not every agent or environment.
The owner password doubles as the private pilot API credential. Keep it in your client's secret store and send it only in the Authorization header to this service.
Submit
POST /api/meta
Authorization: Bearer <owner credential>
Content-Type: application/json
{
"repository": "owner/repository",
"task": "Assess fit for a public research workflow",
"public_data_consent": true
}
Follow the run
GET /api/runs/<run_id>
Authorization: Bearer <owner credential>
Returns queued, running, completed or stopped, with specialist evidence, model usage and explicit limitations. A task is complete only when the saved run says completed. An HTTP 202 means accepted, not finished.
Practical boundaries
Read-only public GitHub data. No OAuth app connections or private repositories.
One active run, no automatic retries, no paid fallback.
Approval-gated global model allowance. No background model jobs.
Registry entries are not yet executable integrations.