# agent-harness

Databricks tested coding agents on day-to-day tasks from its multi-million-line codebase: same model, same reasoning strength, only the harness changed—and on some cases the per-task cost differed by more than 2× with quality held constant. Pi sent about one-third as much context per turn as the comparison baseline. Model and harness both have to count toward total engineering cost.

Kun Chen mapped his daily agent fleet: one "firstmate" managing persistent "secondmates," which spin up disposable "crewmates" per task. Each crewmate gets routed to whichever model is the best deal for the job. gu-log runs its own translation pipeline on the exact same logic.

Without guardrails like CI, the pre-commit gate, the tribunal, and the validator, how bad do AI-written articles get? gu-log has 500+, and the answer needs no imagination — this post, SD-26, is the specimen: it passed every score and still read very AI. The story of a picky editorial desk.

When an agent breaks in production, observability hands you a gorgeous autopsy — every call, latency, and token, but not why it broke or how to fix it. The fix is a loop that runs itself: failure → approved patch → locked-in regression test. Opik is just the example; the point is the loop.

Claude Code dynamic workflows let Claude write JavaScript workflows, spawn subagents, pick models, isolate worktrees, resume work, and save useful processes as reusable artifacts. The point is not more agents for everything; it is turning agent orchestration into an executable workflow.

Cursor moved cursor.com from a headless CMS back to raw code and Markdown. The important part is not just the $260 bill. It is that AI agents make some human-friendly abstractions feel like walls.

An arXiv paper found that inline grep often beats vector retrieval on long-memory conversational QA, while AKBP turns agent memory into a local-first, review-gated, file-backed protocol. Together, they point to the same lesson: agent memory is not a search feature. It is systems engineering.

Whether Claude Code works inside a large codebase is not just about model scores. The real question is whether the team has built rails for the agent: maps, automation, on-demand tools, symbol navigation, internal-system access, and someone to maintain the whole operating setup.

OpenClaw and Hermes are both handing low-level coding-agent execution to Codex app server. This is not just a model switch. It is the agent product stack separating model, execution engine, and chat surface.

Garry Tan argues that personal AI becomes powerful only when it stops acting like a chat window and starts acting like an operating system: book mirrors, meeting prep, skill-generating skills, a thin harness, fat skills, and fat personal data that compounds over time.

A context window is a model's day: how many lessons, messages, tool results, and task events Ryland can experience before sleep, compression, or collapse.

OpenAI DevX's Dominik Kundel says Chronicle means he no longer packages context for AI: one line can sync docs, edit markdown, open a PR, and DM Slack. Nice, but Chronicle's costs are real: screen recording, unencrypted local memories, and prompt-injection risk.

Hermes Agent and OpenClaw shipped big releases the same day. Hermes v0.10.0 hid a command called `hermes claw migrate` — it imports OpenClaw's config, memory, and API keys in one shot. ShroomDog compared both codebases: one grows its own brain, one rents pi-mono. Stay or move?

A 25-person agent platform tore down its engineering pipeline and rebuilt it around one idea: agents are the primary builders. Result: 3-8 prod deploys a day, bad features killed same-day, six-week cycles now land in hours. Harness engineering, applied.

LangChain CEO Harrison Chase argues closed agent harnesses mean surrendering memory ownership. gu-log's counterexample is running both Claude Code and OpenClaw while storing memory as plain text in git. The lock-in is memory format, not harness licensing.

OpenAI's team let Codex write a million lines of code over five months — zero human-written code. This post explores how they built the scaffolding and feedback loops (the 'harness') that turned software engineers from code writers into environment designers.

Everyone's chasing the strongest Model, but the real difference-maker for Agents is the Harness. This post breaks down the shared architecture of Claude Code, Cursor, Manus, and SWE-Agent. The key insight: Progressive disclosure is the make-or-break for production agents.