- Grounding Is Not Proof: 100% Code, 0% Plan Quality
Why a 100% task completion rate does not guarantee plan quality, and how pre-finalization checks prevent hallucinated path claims.
3 min
- Inside the Sandbox: Unconstrained Shell Execution
Exploring why allowing agents arbitrary host shell access leads to non-deterministic failures and why strict Docker isolation is mandatory for production.
2 min
- Targeted Trajectory Repair: Continuations vs Reruns
How provenance-linked continuation child runs inherit parent trajectory history to perform targeted grounding repairs at a 90% reduction in LLM compute costs.
2 min
- The Token Efficiency Curve: Fences vs. Context Dumps
Analyzing how single-cell turn discipline impacts context growth and token costs.
2 min
- Zero Net Access: Hardening LLM Agent Sandboxes
An in-depth analysis of securing autonomous LLM agents using host-only network isolation, Unix Domain Socket proxies, and cryptographically signed audit logs.
3 min
- Strategies, Not Model Ifs
Different models need different recursive-loop policies. The harness separates who from how, binds them optionally, and branches on no model name.
5 min
- RLM Is Not Automatically Token-Efficient
Recursive loops are not token-efficient by default. Compaction plateaued input near 14-16k per call and cut a 25-step run from ~1.04M to ~367k tokens.
5 min
- When Smaller Cells Make Agents Worse
Prompting for smaller REPL cells cut mean cells/turn from 5.0 to 1.5 — but on a short step budget only the multi-block baseline finished.
5 min
Back