Project-Specific Agentic Coding Harness

Problem

A team adopting AI coding tools on a large existing codebase gets a generic assistant: it has not read the project's architecture, does not know which patterns are already proven there, and cannot tell a house convention from an internet average. Its output has to be rewritten, and trust in the tooling collapses before it earns any. Onboarding that toolchain onto a hand-written system people already depend on is the harder half of the problem: the risk is not that agents write no code, it is that they write plausible code nobody reviews properly.

Approach

The harness starts from the codebase rather than from a template. It assesses what is actually there — language, framework, layering, testing approach, and the conventions already in use — and generates from that assessment a tailored set of specialized agents with separated responsibilities: business analyst, architect, engineer, QA, reviewer, and an orchestrator that routes work between them. Alongside the agents it generates project-specific skills, code-review skills, and stack pattern libraries derived from how the codebase already solves problems, so an agent proposes the project's patterns instead of the average of everyone else's. Role separation is a guardrail rather than a metaphor: the identity that implements a change is not the identity that reviews it or signs it off, and the quality gates are deterministic and resumable, so a run either clears the project's own bar or stops where it failed. Two modes are supported — bootstrapping a new project that has no conventions yet, and retrofitting an existing hand-written one, where the conventions have to be read out of the code instead of chosen. Delivery does not end at handover: the harness is tuned with the team as the codebase and its practices move.

Role

Designed and delivered a project-specific agentic coding harness for a production codebase at a national performance-marketing company — the codebase assessment, the agent and skill architecture, the review gates around it, and the ongoing tuning with the team that uses it.

Stack

• Agentic development harness design
• Specialized agent roles and orchestration
• Project-specific skill libraries
• Stack pattern libraries
• Deterministic review and quality gates
• Existing-codebase assessment and onboarding

Outcome

• — Delivered as a paid engagement against a real production codebase, not a demo or a template repository
• — Agents are generated from the codebase's own conventions, so proposals match the project rather than a generic house style
• — Responsibilities are separated across agent roles — the identity that implements a change is never the one that reviews and signs it off
• — Quality gates are deterministic and resumable: a change either clears the project's own bar or the run stops where it failed
• — Works in both directions — bootstrapping a new project and retrofitting an existing hand-written one — and is tuned with the team after handover