Mycelium

Lucy is the squad's executor. It takes a single planned, scoped step and writes the code for it — the only agent with file-write access, and the only one whose work must clear an unconditional compile-and-format gate it cannot talk its way past.

Lucy — Executor agent

Lucy

Executor
Status
Active
Brain
GLM-5.2 (open-weight)

Measured performance

71.6tok/s
Throughput
0.40J/tok
Energy
29W
Power

Qwen3-Coder-30B-A3B-Instruct-8bit · measured 2026-06-18

From the harness

current brain
glm-5.2hosted — open-weight, local planned
last run
2026-07-03 · completed
runs, last 30 days
126 fired · 86 completed

as of 2026-07-03 19:51 UTC · brain read from the harness dispatch map at deploy; runs from the workflow log

Capabilities

  • Code execution

    5/5

    Implements one planned step end to end: edits the files, runs the build, and iterates against the gate until it compiles.

  • Tool use

    4/5

    Drives read, write, edit, shell, and grep tools; a format hint keeps its tool calls in the shape the dispatch loop parses.

  • Scope discipline

    4/5

    Reliable on small, bounded steps of roughly one file; broad multi-file tasks are decomposed before they ever reach it.

Development

  1. done

    Strongest local coder brain

    Swapped to the current coder model after an A/B on a fixed coding suite. The squad runs serially — plan, then code, then verify — so the choice was made on coder quality rather than memory footprint.

  2. done

    Rightsized to a MoE coder

    Swapped from Qwen3-Coder-Next to Qwen3-Coder-30B-A3B - a 30B mixture-of-experts with ~3B active. The larger dense Coder-Next needed ~79GB and stalled on long generations; the MoE recovers coder quality at ~30GB and fits where the big model didn't, without giving up gate pass-rate.

  3. now

    GLM-5.2 in the chair

    Swapped Lucy to GLM-5.2 — a frontier open-weight (MIT) coder. It runs over a hosted API for now, which is the one spot the setup isn't local yet; but open weights mean it's ours to run, and the plan is to bring it home onto a Studio. The swap proves the point of the whole thing: the seat didn't change. Same role, same file-write access, same compile gate it can't talk past. The brain is a dial — the harness is the machine.

  4. done

    A harness fix, not a prompt patch

    Under squad-sized context the coder defaulted to XML-shaped tool calls the loop dropped on the floor. The fix was a format hint in the harness that recovers them — the kind of fix that closes a whole failure class instead of one run.

  5. ahead

    Coder-Next-80B, brought home

    Coder-Next-80B is the stronger coder we want back in the seat — and as of this month it runs on owned hardware: an mmap-elastic loader lets the 80B model sit resident at a fraction of the RAM it used to need, so it fits alongside the rest of the squad. Bench it in the chair, and the coder is fully local again.

  6. ahead

    Wider gated slices

    Carrying larger steps as the gate and the task-scoping around it tighten.