Skip to content

Agent Plane

Peter Pak edited this page Jul 16, 2026 · 8 revisions

Agent Plane

One MCP tool surface, four harnesses, everything recorded. This is the instrument of the thesis: because the tools, database, and printer are identical across harnesses, any difference in behavior is attributable to the harness/model. Design overview: Architecture §B.

The MCP server (plugin/)

TypeScript MCP server, npm @ppak10/agentic-sls-mcp. 14 tools in two groups; full ground rules in plugin/AGENTS.md.

Control (live printer, via Inova API Plugin on :5001):

Tool Effect
printer_status Read-only: layer progress, temperatures, power, positions, override state
recoater_passes_get/set Staged powder delivery within one layer (see Hardware Reference — not N full recoats)
recoater_full_passes_get/set Expand each layer into N full recoat cycles (optionally dry)
layer_overrides_get/set Runtime knob registry: recoater speeds, shake, powder dosing, Z clearance, delays

Knowledge (recorder Postgres, forced read-only pool):

Tool Effect
builds_list / build_get Build catalog and per-build card (linkage, profile params, telemetry summary, ASTM outcomes)
astm_query D638 tensile / D790 flex results by batch/profile/material/specimen
telemetry_summary Thermal tracking error, recoat timeline, laser duty for one build
db_query Single read-only SELECT escape hatch
reference_list/get Curated corpus in the Agentic-SLS-Knowledge dataset submodule (source/): TDS values, Fuse 1 ↔ Inova translation, profile conventions — the submodule pin records which corpus version agents saw

Every *_set call is logged server-side to agent_actions at the MCP choke point — this is how interventions become queryable evidence (and how reflections avoid trusting agent self-reports). It is expected behavior, not something to bypass.

The four harnesses

Claude Code, OpenCode, Codex, Antigravity (agy) — all verified against the same tools. Packaging is one core + thin manifests per harness (Claude plugin marketplace, .opencode/plugins/, Codex config.toml snippet, gemini-extension.json).

Harness quirks worth knowing:

  • codex exec needs --dangerously-bypass-approvals-and-sandbox for MCP (openai/codex#16685).
  • agy reads MCP config only from ~/.gemini/config/mcp_config.json.
  • Claude's built-in Bash/Edit/Write aren't restricted by --allowedTools — the analyst preset denies them explicitly.
  • Evaluate how a harness answered, not just the answer — agy once improvised its own MCP client script and still got the right number.

Presets

  • analyst (default) — read-only tools. Hard-enforced for Claude (including denying built-ins); prompt-level for the others, whose permission flags are all-or-nothing.
  • operator — full control, behind a GUI confirm gate. Intended for zero-shot benchmark runs.

Runner, broker, GUI

  • harness/run.ts — uniform headless driver: npx tsx harness/run.ts <harness> --prompt '…'.
  • harness/server.ts — chat broker on :3100: SSE turns, session resume, preset selection.
  • services/web/ (:5173, docker compose) — Builds registry pages, Agents session browser, right-docked chat panel.

Everything is recorded

Every session/turn lands twice: normalized in Postgres (agent_conversations → agent_sessions → agent_messages; one parser for all four harnesses in harness/store.ts) and raw in the Conversations dataset repo (datasets/Agentic-SLS-Conversations/source/sessions/ — commit it often). Failed runs are data too.

Session rows carry (harness, model, role, build). The comparison design: harnesses bundle models, so OpenCode×Claude vs Claude Code×Claude isolates the harness variable, and OpenCode×{models} isolates the model variable — OpenCode being the strongest local-model harness.


See also: Roadmap for the learning loop that consumes all of this

Clone this wiki locally