•
Agents are writing more and more code
•
They write solid code, until they don’t
•
New bottlenecks
• •Keeping code quality
• •Steering
• •Containing blast radius

•
Platform Teams attempts to keep quality standards for developers
•
Harness Engineering does that for agents

•
Prompt Engineering
•
Context Engineering
•
Harness Engineering

•
Claude Code preview: Feb 25
•
Claude 3.7 Sonnet launch
•
TUI / CLI based agent

•
Opus 4.5
•
Step change in quality
•
Model can prompt itself
•
Improved planning mode

•
The IDE starts losing ground
•
The harness was the main improvement
•
Competition shows up
• •Codex · Opencode · Amp · Pi
• •Gemini CLI · Copilot CLI · Cursor CLI · Droid · Crush
• •Goose · Kiro · Qwen Code · Kimi CLI · Cline CLI
•
It’s not magic
•
The harness matters more than the model

•
It can’t read your files
•
It can’t run your tests
•
It can’t call an API
•
To act in the world, it needs tools

•
One pattern emerged: think, then act
•
Reason → act → observe → repeat
•
The harness runs that loop

•
Non deterministic -> deterministic
•
Code is verifiable
•
Feedback loop

•
You can customize it
•
You can extend it
•
You can build infrastructure for it

•
AGENTS.md
•
A skill
•
A hook
•
An extension

•
Guides steer before the agent acts
•
They make the first try right more often
•
Sensors check after the agent acts
•
They make a second try possible

•
Computational: deterministic
•
Fast, cheap, same answer every time
•
Inferential: non-deterministic
•
Can judge meaning, but slower and can be wrong

•
Observability enables self correction
•
Design agent friendly tools, CLIs/MCPs
•
Favor them over raw text instructions

•
Machine Readable Output
•
Self describing schemas
•
Safety rails against hallucinations
•
Instrument the repository for agents

•
Token Efficiency
•
Reliability
•
Reproducibility
•
Auditable

•
Teams add their own skills and plugins
•
Experts bring the taste
•
Docs live next to the code
•
Your app is a domain-specific harness

•
The bitter lesson: wait for a better model
•
Neuro-symbolic: the model needs code around it

“In the long run, general approaches that scale with available computational power tend to outperform ones based on domain-specific understanding.”

“Building in how we think we think does not work in the long run.”

“Combines neural networks and symbolic AI … to create more robust, more reliable, and more trustworthy AI.”

“Claude Code isn’t better because of scaling. It’s better because it is neurosymbolic.”

“The models get better and better. And so the harness needs to become more and more complicated to allow the model to do more things.”

•
Scaffolding that teaches it to think: the next model eats it
•
Checks that connect it to your world: they stay
•
Don’t teach the model to think. Give it ways to check.

“A good harness should not necessarily aim to fully eliminate human input, but to direct it to where our input is most important.”

An economic and ecological concept describing how individuals, acting strictly in their own self-interest, deplete or spoil a shared resource, ultimately ruining it for everyone

•
Company Scaling Implications
• •Shared standards
• •Tooling
• •Guardrails

•
Fast moving environment
•
Agent Evals and benchmarking
•
Skills versioning and distribution
•
Safely expose company systems for agents
•
Research and Development

•
A model alone can't act. The harness gives it a loop.
•
Guides before it acts, sensors after
•
Same mistake twice? Fix the harness first.
•
There is now a new middle loop
•
Someone needs to own the commons

•
Mastering Harness engineering
• •Enables shipping faster
• •Fewer regressions
• •Lower Organizational drift
•
No one-size-fits-all
•
Start experimenting and find out what works best for your team/org

•
Böckeler, "Harness engineering", martinfowler.com
•
Yao et al., "ReAct", 2022
•
Sutton, "The Bitter Lesson", 2019
•
Marcus, "The biggest advance in AI since the LLM", 2026

•
More on harnesses, agents and loops
•
iurysouza.dev
•
Link and QR on the last slide