Mycelium

Merlin is the squad's deep-research and ingest wizard, with an owl, Archimedes, that folds his context down as he runs so he can read for a hundred turns without losing the thread. Point him at a repository or a question and he maps it, reads the load-bearing files, and writes a cited knowledge state into the squad's shared memory.

Merlin — Deep Research & Ingest agent

Merlin

Deep Research & Ingest
Status
Active
Brain
QUEST-35B-RL (4-bit, MoE)

Measured performance

137.3tok/s
Throughput
0.24J/tok
Energy
32W
Power

QUEST-35B-RL-mlx-4bit · measured 2026-06-18

From the harness

current brain
QUEST-35B-RL-mlx-4bitlocal — oMLX on the Mac
last run
2026-07-03 · completed
runs, last 30 days
4 fired · 2 completed

as of 2026-07-19 00:49 UTC · brain read from the harness dispatch map at deploy; runs from the workflow log

Capabilities

  • Deep research L5

    5/5

    Reads and synthesizes across large repositories and long source sets, running for dozens of turns on its own RL-trained harness.

  • Ingest & condensation L5

    5/5

    Archimedes (a tiny condenser model) folds older context into a compact running state, so a long ingest stays inside the window without dropping what matters.

  • Repository mapping L4

    4/5

    Pulls a repo's full file tree, then reads the load-bearing files across subsystems and cites each path it actually read.

  • Local + remote reach L4

    4/5

    Reads public repos over the GitHub API and private local repos through a scoped file tool that blocks secrets and anything outside the project.

Development

  1. done

    Born on his own harness, served honestly

    Stood up on OSU-NLP's native QUEST harness against our oMLX serving, not a reimplementation. The first runs hallucinated until we found the cause: our server was parsing QUEST's tool calls into a structured field and its thinking into a reasoning field, leaving the content our harness read empty. Reconstructing what QUEST actually emitted turned guessing into grounded research, and a refusal to invent when it couldn't verify.

    Field note: the tool calls we were throwing away
  2. now

    Mapped and ingested the platform and the app

    Read the public Mycelium platform (18 files across the server, routes, SDK, and MCP) over the GitHub API, and the private macOS app (14 Swift files across auth, the platform client, the Cockpit and Desktop faces) through the scoped local-read tool. Both knowledge states are cited and written into the vector memory, retrievable.

  3. ahead

    Ahead

    Cite the local paths he actually read instead of reconstructed URLs, widen the document types he handles, and let him cook for the full hundreds-of-turns runs the harness was built for.