@Michaelzsguo: ds4-agent is cool because it shows how small a coding-agent harness can be (and how close it is with the model is anoth…
Summary
A tweet reflects on how small a coding-agent harness can be, comparing ds4-agent, Pi, and Claude Code, and arguing that the harness—not just the model—is what turns intelligence into software work.
View Cached Full Text
Cached at: 05/23/26, 10:08 AM
ds4-agent is cool because it shows how small a coding-agent harness can be (and how close it is with the model is another story for another day).
Its initial tool surface is tiny: bash, read, write, edit, search, list, more, bash_status, and bash_stop. Depending on how you count bash_status/bash_stop, that is basically 8 or 9 tools.
But this small set is enough to explain what every coding agent harness needs.
Every model needs a good harness. Even DeepSeek is effectively building its own agent harness around the model, not just exposing raw intelligence.
Not coincidentally, the minimalist Pi coding agent by @badlogicgames goes even further: four tools only.
read write edit bash
The model already knows Unix, so bash can cover grep, find, ls, tests, git, and a lot more.
Claude Code is now much more productized, with skills, MCPs, subagents, memory, permissions, hooks, and workflow state. But the legend is that it also started from a very small core: bash and grep.
So what do all coding-agent harnesses have in common?
They all need the same basic loop:
See the workspace. Find relevant code. Edit files. Run commands. Inspect results. Control risk. Repeat.
The model is not the whole agent. The harness is what turns model intelligence into software work.
Similar Articles
@shao__meng: Why do Claude Code, Cursor, Codex, Aider, and Cline exhibit different agent behaviors despite potentially sharing the same underlying models? @addyosmani argues: It's due to the "shell" above the model — the Harness, which includes "prompts, ...
The article discusses how Addy Osmani argues that the performance difference between AI coding agents like Claude Code, Cursor, and Cline stems from their 'Harness'—the layer of prompts, tools, and constraints around the model—rather than the underlying model itself. It details best practices for harness engineering, including hooks, sandboxing, and context management, to bridge the gap between model capability and actual agent performance.
favorite Agentic Coding Harness
The author compares several agentic coding harnesses (Codex CLI, Claude Code, Gemini CLI, OpenCode, Pi) and finds Pi the leanest and best for local models, praising its simplicity and compatibility with Qwen 27B-MXFP8.
@NFTCPS: HarnessX is pretty interesting: an agent architecture that can modify itself. Previously, architectural changes relied entirely on manual tuning. When a new model came out, Anthropic removed the planning steps from Claude Code, and Manus refactored its agents five times in six months, each time simplifying. What to change and when to change it — all decided by humans.
HarnessX introduces a framework for self-evolving AI agent harnesses that treats the runtime harness as a first-class object, enabling automatic adaptation via trace-driven reinforcement learning. It achieves average gains of +14.5% across five benchmarks, with larger improvements for weaker models.
Pi Coding Agent
Pi Coding Agent is a customizable harness for building coding agents, now available on Product Hunt.
Observation: the best agent harness for each model will be from the model developer themselves
A discussion on how AI models perform best with harnesses developed by their own creators, as third-party harnesses may cause underperformance despite strong benchmarks, citing examples like Claude Code for Claude and Codex for GPT.