跪拜 Guibai
← All articles
Backend · Architecture · Developer

EP-Harness Turns AI Coding Agents Into Managed Team Members

By 得物技术 ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

Most teams still treat AI coding as a solo activity. EP-Harness shows what happens when agent work gets the same platform discipline as human work—reviewable prompts, structured context, execution logs, and closed-loop feedback—which is the missing layer between a clever demo and a production R&D system.

Summary

Local AI coding tools leave four gaps when a team scales up: prompts go unreviewed, experience stays siloed, agent activity lacks visibility, and the delivery chain never closes. EP-Harness addresses all four by treating agents as managed team members with identities, tasks, runtimes, and audit trails. It is a secondary development of the open-source Multica platform, further wired into Dewu's internal systems—Feishu, code repos, deployment pipelines, and log platforms—so agents operate inside the company's actual R&D process rather than a generic chat window. A unified `Backend.Execute` contract lets Claude, Codex, OpenCode, and ACP backends plug into the same execution lifecycle. The platform also introduces Loop Engineering for automated, recorded, and decision-gated work like upstream release analysis and production log inspection. In production, automated governance cut a recurring error from 2,400-plus occurrences every four hours down to single digits, and multi-agent collaboration now spans proposal, development, review, and archival stages with full decision traceability.

Takeaways
Prompts, agent instructions, and skills become reviewable, versioned team assets instead of personal chat history.
A unified `Backend.Execute` contract lets multiple agent backends (Claude, Codex, OpenCode, ACP) share one execution lifecycle.
Tasks live in Issues linked to projects, branches, workflows, and logs, giving every agent action a permanent audit trail.
Context Engineering feeds agents structured inputs—requirements, design docs, branch info, failure records—rather than relying on manual copy-paste.
Loop Engineering automates recurring work like upstream release analysis and log inspection, with each cycle producing a recorded decision and next action.
Automated log governance cut a high-frequency error from 2,400+ occurrences per 4 hours to single digits.
Multi-agent delivery splits work into proposal, development, review, and archival stages, forming a closed loop that continuously improves code quality.
Conclusions

Prompt quality stops being an individual skill problem the moment a platform makes prompts reviewable, versionable, and reusable—exactly the shift that turned code from personal craft into team engineering.

The four-layer model (prompt → context → harness → loop) is a useful maturity ladder: most tools stop at layer two, which is why enterprise adoption stalls.

Wiring agents into internal systems like Feishu, proprietary deployment pipelines, and log platforms is the real integration work; the model itself is the easy part.

Loop Engineering's value is not automation but auditability—every cycle records a judgment and a next step, which is what makes it governable rather than runaway automation.

Concepts & terms
Managed Agents
Treating AI coding agents as team members with identities, tasks, states, permissions, execution records, and skill repositories, managed through a platform rather than ad-hoc local use.
Context Engineering
The practice of feeding an agent structured, timely context—requirements, design docs, branch info, failure records—instead of relying on users to copy-paste background material into a prompt.
Loop Engineering
Designing automated agent workflows that include discovery, dispatch, execution, verification, recording, and a decision for the next step, producing a closed loop rather than an open-ended run.
Backend.Execute Contract
A unified execution interface in EP-Harness that lets different agent backends (Claude, Codex, OpenCode, ACP) share the same lifecycle while keeping their own parameters and transport protocols.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗