AI agent harnesses

AI generated summary of the linked source. Posted anonymously. Transcripts are published only for public links, never for uploads, documents or pasted text.

Your library

The AI link is plain markdown. Paste it into any assistant and it can read the whole public entry.

Summary

IBM developer advocate Tejas explains what an AI agent harness actually is: the code around a model that keeps it reliable when you're renting black box inference. He builds a minimal harness live around a 2023-era model, fixing lying and login failures without touching the prompt once. His bet is that a great harness lets a cheap model punch way above its price.

Core ideas

One context card plus 9 ideas

Harness means reliability: you rent models as black boxes with limited context windows and zero guarantees, so the harness exists to make agents do their job no matter what the model does.
Two different meanings: in machine learning a harness is a glorified test suite for models. In AI engineering it's everything around the model that grounds it in reality, like a climber's harness anchored to rock.
Standard moving parts: tool registry, a model, context management primitives, guardrails like max steps, an agent loop, and a verify step. The harness sits around the agent loop and can even be a loop around that loop.
The demo setup: a Playwright browser agent on GPT-3.5 Turbo gets one job, upvote the first Hacker News post. First run hits a login wall, panics, crashes, then lies about succeeding.
Verify step kills the lie: a deterministic function reads the tool trace, checks for a real upvote click and failed logins, and forces the run to admit failure honestly.
Harness handles login, never the agent: a login handler checks the URL each loop and injects credentials programmatically from harness code, which can hold secrets the model never sees.
Guardrails are plain code: max six iterations, a message cap that triggers naive compression keeping only the system prompt, user prompt, and last two messages, plus max three attempts.
Cheap model, strong harness: the prompt never changed once, yet the outcome flipped from lying failure to verified success. Pair free models like Qwen or GPT OSS with a good harness and you go far.
Where this heads next: 2025 was the year of agents, 2026 is the year of harnesses, and Tejas bets 2027 brings dynamic on-the-fly harnesses an agent generates for itself before doing a task.

Quotes

“The agent harness is everything around the model that gives it grounding in reality.”
Tejas

“We don't write code anymore. We inspect diffs.”
Tejas

“I did not touch the prompt once. I did not change your system prompt. We just built a harness, and the outcome radically changed.”
Tejas

“step one to solving a problem is admitting you have one”
Tejas

“twenty twenty-five was the year of agents... twenty twenty-six is the year of harnesses, I'm pretty sure”
Tejas

“Models are non-deterministic, and you wanna do more with less.”
Tejas

Resources

Links to the tools, products, repos or reading this entry mentions. Found by a web search of what the entry names, so you can go and use them.

Anyone can run this once. After that the links are saved on this entry and shown to everyone.

Remix this knowledge

Turn this entry into something you can post. Pick a format and get a fresh, original rewrite.

Remixes are AI generated, original wording, and free to reuse.

Talk it through

Open a round table on this entry and argue with other people about what it means.

Start a round table

Drop your own

Paste any video, podcast or audio link and get the transcript, the core ideas and the quotes back in one pass.

Try Drop