Archives

Why Build an Agent Harness?

Zero-Spend Concurrency Auditing for Agent Sandboxes

Catching Hallucinated Paths in RLM Trajectories

Grounding Is Not Proof: 100% Code, 0% Plan Quality

Inside the Sandbox: Unconstrained Shell Execution

The Security Cost of Autonomy: Auditing 29 Runs

From REPL Loops to Swarms: Leasing Sandboxed Containers

Targeted Trajectory Repair: Continuations vs Reruns

The Token Efficiency Curve: Fences vs. Context Dumps

Zero Net Access: Hardening LLM Agent Sandboxes

A Plan Can Validate and Still Be Unsafe to Implement

Benchmarking Agent Systems Beyond Did It Finish?

Designing Bounded Repair Loops for Agent Plans

Experiments Should Be First-Class Product Artifacts

Grounding Is Necessary, Not Sufficient

What a Two-for-Two Agent Result Actually Proves

Strategies, Not Model Ifs

RLM Is Not Automatically Token-Efficient

When Smaller Cells Make Agents Worse