AI · Mike Sulsenti

Layer 01

Benchmarks

Private tasks, published results

B-01

Code Handoff Bench

How well a lead model briefs coding subagents, and how well those subagents write hard, production-grade code from the brief. Each task is one hard file from a real codebase. Entries are graded by a hidden test suite, an anonymous council of judges from different vendors, cost and time from each harness's own logs, and speed against a reference.

Runs: –
Entries: –
Tasks: –
Leads: –

Open results Repository ↗

Top overall picks · all taskscouncil score / 10

Loading results…

Ranked by value: council score, less 0.2 per doubling of cost and 0.5 per point of findings, among entries that pass the hidden suite.

Layer 02

Setup

One interfaceT3 Code → agents

T3 Code is the one window I work in. Every coding agent runs behind it, so threads, diffs and reviews look the same whichever agent is doing the work.

Every devicea web around the hub

Every machine with a screen runs the T3 Code app and can run agents itself, or send work to any of the others. The headless Agent workstation is the hub that carries most of it, and it hands overflow to the Intel Mini PC. The servers behind all this are on the homelab map.

Layer 03

Models

What each one does for me, and at what effort

Model	Roles	Effort

Layer 04

The High Council

A review skill · 3 seats, 3 labs

The High Council is a review skill I run from T3 Code. It hands a diff, branch, pull request or design doc to three models from different labs. Each one reviews it alone, then they cross-check each other's findings, argue out the disagreements and score the work. The round repeats until the council is satisfied.

01Review
02Cross-check
03Deliberate
04Score
05Repeat until satisfied

Layer 05

Lab standings

My ranking of how well each lab is doing right now. Opinion, not a benchmark.