Part 2 · Coding agents

Coding agents
& the coding harness

What sits around the model — and why a better harness leads to better code.

The landscape

Some coding agents

Hermes agentHermes agent
opencodeopencode
Claude CodeClaude Code
GitHub Copilot CLIGitHub Copilot CLI
CodexCodex
… and many more
The landscape

Some coding agents — what do they cost?

freeHermes agentHermes agent
open sourceopencodeopencode
€ € €Claude CodeClaude Code
€ € €GitHub Copilot CLIGitHub Copilot CLI
€ € €CodexCodex
Open-source agents run on your own models — the commercial ones bill per use.
The big picture

Coding agent = model + “coding harness”

MAP

The five building blocks of a coding harness — we’ll walk through each one.

1 · AI model

Coding agent = thin client

The harness picks the model. Swap the brain without changing the tool.

Coding harness
select AI model
API key (e.g. Gemini)
Inference provider
Local model (e.g. Gemma 3)

Inference providers are e.g. Nebius, RunPod, Hugging Face Hub, etc.

Coding agent = model + “coding harness”

MAP
2 · Skills & MCP

Give better instructions with Skills

  • SKILL.md is an open standard by Anthropic.
  • It lets you load a file that instructs the LLM what to do.
  • Most often used with Claude…
  • …but it’s open source, so it runs with any coding agent.

Find skills at officialskills.sh

---
name: my-pptx-style
description: > Use this skill whenever
  creating or editing PowerPoint slides.
  Encodes the user's personal style.
---

# My PowerPoint Style Guide

## Design System
— Font: Aptos
— Text: dark navy on white
— Primary #0D2D6B · Accent #1565A7
2 · Skills & MCP

What is MCP?

MCP (Model Context Protocol) is an open-source standard for connecting AI agents with external systems.

Using MCP, agents can connect to data sources, tools and workflows.

Kind of like a specialised API language for agents!

MCP host connects to MCP server, which exposes APIs, code and databases as tools
2 · Skills & MCP

MCP vs REST

REST is the right choice when…

  • You’re building a service other developers or apps will consume.
  • You want standard HTTP tooling — caching, auth, rate limiting, load balancing.
  • The consumer just needs data, not reasoning.

MCP is best when…

  • An LLM needs to decide which tool to call from a natural-language request.
  • You want one agent to expose capabilities to another agent.
  • The set of available tools isn’t known at code-write time.

Coding agent = model + “coding harness”

MAP
3 · Prompt & context engineering

Prompt & context engineering

Prompt engineering: write better & longer prompts.

Context engineering: give the model the right context documents:

  • Architecture
  • Style guides
  • Patterns
  • Master data

Coding agent = model + “coding harness”

MAP
4 · Memory

Three tiers of memory

The coding harness uses three distinct kinds of “memory” — only one of them actually learns.

🔁

Session state

Short-term · local
  • Everything from the current session.
📝

Instructions

Durable · authored
  • AGENTS.md and .github/instructions/
  • Used from Skills
  • Deterministic config, not learned
  • Settings as SDD / TDD
🧠

Memory

Learned · hosted
  • Captured automatically as it works
  • Repo facts + user preferences
  • Validated; auto-expires after 28 days
  • Shared across CLI, cloud agent, review

Coding agent = model + “coding harness”

MAP
5 · Sub-agents

Multi-agent systems

Example: a city-noise report pipeline — three agents, one task.

Task: “Measure background noise at 5 Amsterdam intersections and produce an ISO 9614 compliance report.”
1
Planner
Generates high-level tasks
Responsibilities
  • Decompose goal into subtasks
  • Assign locations & priorities
  • Set success criteria per task
  • Sequence tasks & timing
Output: ordered task list with criteria
2
Worker
Executes the tasks
Responsibilities
  • Call measurement API per location
  • Apply background-subtraction formula
  • Retrieve noise ontology via MCP
  • Write sections of the ISO report
Output: measurements + draft report
3
Verifier
Controls every result
Responsibilities
  • Check margin ≥ 3 dB per measurement
  • Validate ISO 9614 format compliance
  • Flag uncertain results for re-run
  • Approve or reject each section
Output: verified report or rejection + reason

Pattern: Generator–Verifier nested in Orchestrator–Subagent · Verifier = vier-ogen-principe — no result passes without independent review.
Ref: Feng, McDonald & Zhang (2025). Levels of Autonomy for AI Agents. arXiv:2506.12469

Wrap-up

Summary

We discussed the five most important elements of a “coding harness”:

AI model Skills.md & MCP Prompt & context engineering Memory Sub-agents

This presentation was just a first introduction to the subject.

Better coding harnesses will lead to better code. In the second half of 2026 we’ll see lots of improvements in coding harnesses. Watch this space.

1 / 16
Use ← / → arrow keys, or click the dots