Notes on agents, software, and work.
Systems / August 26, 2026 / 10 min read
Human in the-loop vs human on the-loop: where control sits
HITL pauses selected actions before execution; HOTL supervises more autonomous execution, while undo cost determines where a human gate belongs.
Systems / August 25, 2026 / 10 min read
AI agent frameworks open-source: seven workload fits
Open-source AI agent frameworks fit different workloads; compare eight choices by execution type, use case, license, and GitHub popularity.
Systems / August 16, 2026 / 10 min read
How to choose an open-source agentic AI framework by its control contract
An open-source agentic AI framework earns its place through durable state, scoped tool authority, approval gates, tracing, evals, and crash recovery.
Systems / August 16, 2026 / 10 min read
Best open source agent harness: choose by control, recovery, and workload
Choose an open-source agent harness by context load, carried judgment, reversibility, state, approvals, recovery, maintenance, and operating cost.
Systems / August 16, 2026 / 11 min read
How to build an agent harness backward from failure
Build an agent harness from failure modes outward, with bounded authority, durable state, independent checks, monitoring, and recovery.
Systems / August 16, 2026 / 10 min read
Harness engineering framework: six control responsibilities for reliable agents
A harness engineering framework assigns context, execution, permissions, evaluation, recovery, and tracing to explicit controls.
Systems / August 16, 2026 / 10 min read
AI delegation: how much authority should an agent get?
AI delegation works when each action has a clear authority boundary, proof of completion, stop condition, and human approval point.
Systems / August 16, 2026 / 10 min read
AI agents vs agentic AI: where the acting unit ends and the system begins
AI agents perform tasks for users; agentic AI shifts the unit to long-horizon work, with orchestration and controls shaping the system.
Systems / August 16, 2026 / 10 min read
Agentic AI vs LLMs: where the model ends and the system begins
LLMs generate language; agentic systems add memory, tools, execution, controls, and recovery around the model to complete multi-step work.