ECC

The ECC GitHub repository is a performance-optimisation system for something most teams have not started treating as a product yet: the agent harness — the layer of tooling, memory and instructions that surrounds a model in products like Claude Code, Codex, OpenCode and Cursor. ECC, by affaan-m, starts from a simple observation: the same model produces wildly different results depending on the harness around it, so the harness deserves engineering. With more than 291,000 stars and 43,000 forks, the ECC GitHub project has become a reference implementation for teams who want to squeeze reliability, not just speed, out of their coding agents. The official home of the project is the affaan-m/ECC repository on GitHub, where you can browse the source code, releases and issue tracker.

What it is

ECC is a JavaScript system, released under the MIT licence, that bundles four kinds of optimisation into one installable layer. Skills give the agent reusable, task-specific instructions. Instincts and memory let the harness retain what worked — project conventions, past mistakes, preferred patterns — across sessions instead of starting cold every morning. Security checks watch the agent’s file and command activity for the failure modes that matter when software writes software. And a research-first workflow forces the agent to look before it leaps: read the codebase and the documentation before proposing changes. The unifying idea is operational — treat the harness the way you would treat production infrastructure, with attention to what it does when nobody is watching.

Key features

• Skills system — reusable instruction packs the agent loads for specific jobs, shareable across a whole team.
• Instincts and persistent memory — the harness learns project conventions and retains them between sessions instead of starting cold.
• Security checks — guardrails around agent actions, aimed at the risks unique to autonomous code changes.
• Research-first development — enforced read-before-write behaviour that cuts hallucinated APIs and wrong assumptions.
• Harness-agnostic — built for Claude Code, Codex, OpenCode, Cursor and beyond, so the investment is not tied to one vendor.
• Performance focus — the goal is measurably better output from the same underlying models, not new models.

Who it is for

Teams who have moved past experimenting with a single agent and now run agents as part of their delivery pipeline — and have the scars to show for it: inconsistent output between sessions, repeated mistakes, review bottlenecks. Platform and DevEx engineers will recognise the shape immediately: ECC is to agent harnesses what a good CI setup is to builds. Agencies standardising agent behaviour across many client projects are the other natural audience. It is probably overkill for a solo developer running occasional one-off prompts, where the harness is thin by nature and there is little process to optimise.

Getting started

Start from the ECC GitHub repository: clone it, follow the install path for your harness — the project documents setups for Claude Code, Codex, OpenCode and Cursor — and enable the pieces incrementally: skills first, then memory, then the security checks, so you can feel what each layer changes. Baseline matters here. Before installing, save a few representative agent outputs on real tasks; after a week with ECC, run the same tasks and compare review effort, not just speed. Teams that skip the baseline end up arguing from vibes about a system whose entire point is measurable improvement.

Our take

ECC is betting on the right layer. Models will keep improving and swapping; the harness — your skills, memory, checks and workflows — is the part you own, and it compounds. The honest caveats: harness engineering is real engineering, with its own maintenance burden, and an over-configured harness can make agent behaviour harder to predict, not easier, so adopt in layers and measure as you go. There is also a maturity question — the field’s conventions are months old, so expect the vocabulary to keep shifting under you. Even so, for any team running agents in production, the ECC GitHub repository is the most complete statement of the harness-first philosophy we have seen, and its star count suggests the industry agrees.