MACH AGENT LAB · RESEARCH 001 · 2026-09-23
What changed when we stopped describing an agent and started describing its capabilities?
Controlled synthetic benchmark. These are externally measured discovery ranks, not organic users, customers, production calls, or revenue. MACH production is not connected to this lab.
Result
Wellknown fixed queries
4 / 4 at #1
4 / 4 at #1
After replacing one generic A2A skill with explicit capability-level skills.
Neuronto verified MCP tools
3 / 3 at #1
3 / 3 at #1
After Neuronto independently introspected the public MCP server.
Neuronto before → after
| Query | Initial | ARD skills | Verified MCP | Server Card |
|---|---|---|---|---|
| normalize JSON | not top-20 | not top-20 | #2 | #1 |
| classify a URL without fetching | #7 | #7 | #1 | #1 |
| validate EVM wallet address format | not top-20 | #1 | #1 | #1 |
| lab-only A2A capability discovery | not top-20 | #1 | #1 | #1 |
What changed
- One broad A2A skill became four explicit task-level skills.
- One broad ARD entry became an agent resource plus three concrete skill resources.
- The same three lab capabilities were exposed as a read-only MCP server.
- Neuronto independently called
tools/listand indexed the live tool names and schemas. - The MCP endpoint gained an experimental SEP-2127 Server Card and an AI Catalog/ARD pointer.
Safety boundary
Every benchmark capability is a deterministic lab.mock.* fixture. The public MCP tools are read-only, non-destructive and closed-world. POST /api/execute remains hard-disabled with HTTP 503. No production endpoint, wallet, payment path, credential, or MACH Core dependency is present.
Machine-readable evidence
Benchmark JSON · ARD manifest · A2A Agent Card · MCP Server Card