Observed arrival · 2026-11-03
Agent Headroom measures how many coding agents a machine can sustain
An Apache-2.0 benchmark suite replays recorded coding-agent sessions on local hardware to measure capacity across memory, speed, concurrency, models, and context.
- For
- Developers sizing local coding-agent fleets
- Worth noticing
- The runner refuses to start if the machine has changed since planning, and result files retain probe and memory-trace evidence.
Field notes
Before a run, a probe reads device and host memory, topology, bandwidth, and storage; the planner chooses model, quantization, context, and placement that fit. The runner pins server flags and replays recorded sessions, with probe, plan, command line, and memory trace stored alongside each result. Displayed examples include a 192 GB Ryzen host reaching two live-tool agents before its host pool fills at four, and a GB10 keeping three models resident.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue