Article · 2026-07-13 · 6 moments

🎙️ How I AI: GPT-5.6 review, How a solo builder runs 24/7 local AI, and What an agent harness is and how to build one

Your weekly listens from How I AI, part of the Lenny’s Podcast Network ✦ AI generated

Definition◆

A harness is simply code you write to make an AI agent more effective at a specific job.

Claire demystifies the term 'harness,' defining it as ordinary code wrapped around an AI agent, noting that even complex tools like Cursor and Claude Code are just elaborate harnesses.

transcript

Claire: The term has taken on an almost mythical quality in engineering circles, but I strip it down in this episode: a harness is code you write to make an AI agent more effective at a specific job. Cursor is a complex harness. Claude Code is a complex harness. Yours can be eight files and a terminal UI.

extends · 2gives example · 1supports · 1

Anecdote

Claude Fable gets too locked into its own frameworks and refuses to reconsider, while GPT-5.6 Sol is willing to adapt and fix problems pragmatically.

Claire recounts a debugging standoff where Fable insisted a broken tool-calling loop was a model limitation, while switching to Codex let Sol fix the same issue on Sonnet 5 in one shot.

transcript

Claire: I had a hardened tool-calling loop in my prototyping product that only GPT-5.5 could run. Fable insisted it was a model problem and refused to budge. The moment I switched to Codex and told it to just fix it, Sol got Sonnet 5 working in one shot.

rebuts · 2supports · 1

Data◆

GPT-5.6 Sol is the most practically effective model for real product work, even though Claude Fable may be theoretically smarter.

Claire's five-category benchmark across PRDs, prototypes, wireframes, debugging, and agentic voice found Sol scoring highest on 'taste,' making it her new daily driver despite Fable's raw intelligence edge.

transcript

Claire: I ran a five-category benchmark across PRDs, prototypes, wireframes, debugging, and agentic voice, and Sol had the highest taste score by a significant margin on the 70% Claire/30% machine split. That gap between 'hyper-intelligent' and 'actually ships' is real, and for product work Sol wins.

rebuts · 1supports · 1

Mechanism

In a hybrid AI workflow, cheap local models should handle high-volume scanning while expensive frontier models like Claude Code handle judgment calls.

Alex Finn's security scanning setup pairs a cheap local model that scans code every 20 minutes with Claude Code, which reviews findings once a day to judge what's real.

transcript

Alex Finn: A local model (GLM 5.2) scans code every 20 minutes and dumps findings into a Markdown file. Claude Code checks that report once a day and decides what's real and worth fixing. The local model does the volume work cheaply; the frontier model does the judgment work precisely. Trying to run Claude Code every 20 minutes instead would cost thousands a month.

gives example · 2

Claim◆

The case for local AI hardware isn't near-term ROI—it's the unlimited inference that makes running agents 24/7 economically viable.

Alex Finn argues that buying expensive local hardware only makes sense once you factor in running agents continuously, since cloud API costs would otherwise spiral.

transcript

Alex Finn: The math on a $10,000 Mac Studio vs. a $20 ChatGPT subscription only looks crazy until you run an agent 24/7. At that scale, cloud APIs get expensive fast, and local models running around the clock open use cases that simply aren't economically viable otherwise.

extends · 1supports · 1

Claim

You should build a custom harness when a workflow needs the same setup and produces the same outcomes every time, combining deterministic steps with non-deterministic AI judgment.

Claire explains the trigger for building a harness: a job that's partly fixed process and partly AI judgment, using her Sentry bug-triage harness as the qualifying example.

transcript

Claire: The trigger is recognizing a job that is partly deterministic (defined steps, defined tools) and partly non-deterministic (the AI figures out root causes and writes the report). Sentry bug triage qualified because every investigation follows the same evidence-gathering process and ends with the same artifact bundle.

provides context · 1supports · 1

Highlight slides

Local AI Hardware Pays Off at Scale✦ from: The case for local AI hardware isn't near-term ROI—it's the unlimited inference that makes running agents 24/7 economically viable.What a 'Harness' Really Is✦ from: A harness is simply code you write to make an AI agent more effective at a specific job.Sol Wins on Practical Product Work✦ from: GPT-5.6 Sol is the most practically effective model for real product work, even though Claude Fable may be theoretically smarter.Why 24/7 Changes the Calculus✦ from: The case for local AI hardware isn't near-term ROI—it's the unlimited inference that makes running agents 24/7 economically viable.Simple Idea, Any Scale✦ from: A harness is simply code you write to make an AI agent more effective at a specific job.Taste Beats Raw Intelligence✦ from: GPT-5.6 Sol is the most practically effective model for real product work, even though Claude Fable may be theoretically smarter.

Related episodes

Article · Latent Space

[AINews] OpenAI launches GPT 5.6 Sol/Terra/Luna, Codex becomes ChatGPT superapp

→

Article · Simon Willison's Weblog

The new GPT-5.6 family: Luna, Terra, Sol

→

Video · The Cognitive Revolution

Intelligence on the Edge: Liquid AI's Ramin Hasani on the Search for Device-Native Foundation Models

→