Skip to content

Agent Mesh and Tooling Standard

Lucas YourLastName edited this page Jul 31, 2026 · 3 revisions

Agent Mesh And Tooling Standard

This page documents the generic SYSTEMX standard for Agent 0, subagent lanes, browser automation, MCP tooling, desktop control, external service connectors, and recovery paths.

The public template keeps this vendor-neutral. Project-specific providers, customer workflows, private dashboards, and proprietary adapter logic belong in the project using the template, not in the template itself.

Canonical Source

The canonical files live under .SYSTEMX/AI/:

File Purpose
.SYSTEMX/AI/AGENT-MESH-STANDARD.md Agent 0, subagent lanes, message bus, checkpoint, memory, and archive rules.
.SYSTEMX/AI/TOOLCALLING-AND-BROWSER-AUTOMATION.md Playwright, Chrome DevTools MCP, local browser control, and desktop fallback rules.
.SYSTEMX/AI/EXTERNAL-SERVICE-CONNECTOR-STANDARD.md Generic adapter pattern for outside services, SDKs, CLIs, dashboards, and webhooks.
.SYSTEMX/AI/RECOVERY-PLAYBOOK.md Popup, permission, auth, port, process, and apply/blocker recovery paths.
.SYSTEMX/AI/agent-mesh.schema.json Minimal machine-readable schema for durable agent messages.

Run the standard check locally:

npm run ai:standard:check

Agent 0 Pattern

Agent 0 is the coordinator. It owns the mission, chooses lanes, reviews evidence, decides when to archive, and asks the operator for credentials, production approval, paid-service actions, or ambiguous account decisions.

Subagents are scoped lanes, not hidden authority. Use lanes such as research, code, test, browser, desktop, connector, security, docs, and release.

Every durable checkpoint should include mission ID, wave ID, lane, sender, receiver, status, summary, evidence, blockers, next action, and creation time.

Browser And MCP Loop

Use this local-first route:

npm run dev
npm run browser:install
npm run browser:codegen
npm run mcp:chrome

Use Playwright for repeatable browser flows and Chrome DevTools MCP for live DOM, console, network, and screenshot evidence. Keep these pointed at localhost, emulators, or staging by default.

Apple Silicon And Desktop Automation

On macOS Apple Silicon, automation must account for app permissions, Keychain prompts, GUI shell PATH, browser auth windows, and Accessibility API shape.

The standard is:

  • prefer browser-first auth when a native auth presenter stalls
  • make local UI labels bounded and meaningful for screen tools
  • avoid recursive accessibility dumps
  • stop and replace only the process owned by this repo/session
  • keep secrets in OS keychain or ignored env files
  • never click through paid, destructive, or permission-escalation prompts

External Connector Pattern

Treat every outside service as a connector with mode, auth method, allowed read-only actions, write actions requiring preflight, secret names only, webhook events, evidence expectations, and deny reasons.

This lets projects add vendor SDKs, dashboards, and APIs without teaching the public template private business logic.

Recovery Ladder

When automation hits a popup, stalled apply operation, blocked browser flow, or unknown prompt:

  1. Capture URL, visible text, screenshot, command, and logs.
  2. Classify the blocker.
  3. Try the lowest-risk alternate route.
  4. Record a dead-letter entry if the lane cannot continue.
  5. Ask the operator when the next step involves credentials, production state, paid services, or unclear account ownership.

Dead-letter entries are part of the standard. They prevent Agent 0 and future subagents from wasting tokens rediscovering the same blocker.

Clone this wiki locally