Version-pinned
Every claim names the version, commit and date examined.
Independent research into agent systems
Models reason.Harnesses act.
Harness the Agents examines the systems that turn models into agents: loops, tools, memory, sandboxes, control planes, and the evidence for how they actually behave. We read the source, pin the version, publish the trace.
The publication
Five desks, one evidence standard.
Mechanism first, organisational implication second. Every desk works to the same rule: claims you can check.
- 01
Inside the Harness
Version-pinned teardowns of agent systems: the loop, tools, memory, permissions and execution boundaries, examined in the source at a named version.
- 02
Above the Harness
Enterprise control architecture for agent systems: identity, policy, approvals, budgets, observability and accountability that no single harness provides.
- 03
Harness Test
Reproducible controlled experiments: same task, pinned versions, published prompts, environments, traces and limitations.
- 04
Unharnessed
Evidence-led analysis of agent system failures: what broke, the mechanism behind it, the blast radius and the mitigation.
- 05
Agent Field Notes
A weekly intelligence briefing: what shipped and what it means for the people running these systems.
Featured
Inside
DeepSeek Harness Isn't a Claude Code Killer. It's Something More Interesting.
DeepSeek's open-source agent harness passed 168,000 GitHub stars in its first week. I installed it, read the repo, and ran it myself. Here is what holds up.
The rules of evidence
Source-led
Vendor statements are claims, not evidence. Code and traces come first.
Limits on the record
Uncertainty, failed tests and reproduction gaps are published too.