Breaking

DenisSergeevitch Releases Provider-Neutral Agent Skill for Harness Design

September 11, 2026 at 06:39 EDT

AI Agents
Open Source

DenisSergeevitch has published an open-source, provider-neutral Agent Skill that turns agent-harness engineering into a reusable workflow for MVP design, audits, refactoring, and explanation. The agents-best-practices project is aimed at Codex, Claude Code, and other Agent-Skill-compatible runtimes, while stopping short of being an execution framework. The project documentation describes it as a portable reference for systems in which the model is only one part of the runtime.

SEPT. 11, 2026 · OPEN-SOURCE TOOLING

A playbook for safer agent harnesses

The agents-best-practices skill packages provider-neutral guidance for MVP design, audits, refactors, permissions, memory, and evaluation.

What is inside

1.7.0

skill version

reference files

2,281

stars

205

forks

Repository anatomy

One entry file routes agents into a deeper reference library.

1entry skill

21reference files

Size is proportional: one entry point, 21 focused references.

How the loop is supposed to work

1 · Build

Context

→

2 · Propose

Model action

→

3 · Gate

Validate and approve

→

4 · Observe

Record result

The practical boundary

The skill improves the design brief; the host still owns runtime enforcement.

Useful layer

MVP boundaries, typed tools, approvals, budgets, context, and evals.

Still required

Application code, sandboxing, authorization, traces, tests, and provider integration.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.