Anchoring AI — Post 12

The Agent Harness

Putting gitagent.sh to work on a real codebase — from task prompt to pull request

Status: Placeholder. This post is planned. The outline and key concepts below describe what it will cover.


What This Post Covers

Post 11 surveyed the agent framework landscape — GitAgent-style tools, LangGraph, Azure AI Foundry Agents — and explained why MCP-first architecture ages well regardless of which orchestration layer wins. This post stops surveying and starts doing.

gitagent.sh is an agent harness: a runtime that gives an AI agent a task, a repo, and the autonomy to execute across both. The distinction from a coding assistant is fundamental — you’re not prompting interactively and reviewing each step. You’re defining a task, handing it off, and getting back a result (code, a PR, a report). The agent plans, reads, writes, and commits without you in the loop.

This post walks through connecting gitagent.sh to this contact-center-ai codebase: what the harness is, how it’s configured, what tasks are a good fit (and what aren’t), and what actually happens when you point it at a real backlog item.


Key Concepts


Planned Outline

  1. From assistant to harness — the conceptual shift: what you give up (interactive control) and what you gain (delegation at task granularity); when that trade-off makes sense

  2. What gitagent.sh is — the harness model: task file → agent execution → diff/PR output; what the agent can do (read files, run shell commands, write code, commit); what it can’t (make architectural decisions without context, handle ambiguous requirements)

  3. Setting it up with this repo — configuration walkthrough: pointing gitagent.sh at the contact-center-ai repo; CLAUDE.md as the primary grounding document; what environment the agent runs in

  4. Choosing a first task — picking something from the actual backlog:

    • Fix the query_csat() category filter bug (known bug from Post 3, small scope, clear success criteria)
    • Or: implement the S3 deduplication guard from Post 5
    • Criteria: bounded, testable, consequence of getting it wrong is low
  5. Running the task — what the execution looks like; what the agent reads first; how it plans; what it produces

  6. Reviewing the output — reading the diff; what the agent got right; what needed adjustment; the human judgement calls that belong to the reviewer, not the agent

  7. The MCP-as-data-source angle — an agent implementing a feature that queries call data through the MCP server to inform its implementation choices; the loop closes

  8. What this changes — how the development workflow shifts when some tasks go to an agent; what stays human; the right mental model for agent harnesses as a team tool rather than a solo accelerator


Code Changes for This Post


Outstanding Questions / TBD