Ex-Google engineer reveals how to build AI agent loops, harnesses, LLM ops, and evals in 19 minutes. Trace → evaluate → diagnose → fix → ship → repeat. That loop is how agents self-improve over time. Agentic loops + harness + memory + evals - that’s the senior engineer... stack. This is better than $500 paid courses on the same topic, explained in under 20 minutes. Watch it, then save the framework below.show more

You did not ask me for my permission to download and upload my video. Please delete and quote my original post. Thanks. - -

Trace → evaluate → diagnose → fix → ship → repeat by the way, I've already seen a similar connection

Yeah bro. That’s basically senior level engineering stack right now

Sharing this is very helpful.

watching it right away codez

I’m sure you will like it Roan !

nice sharing!

Thanks bro. Are you building self improving systems yourself ? Loops, harness ?

This is good but just beware that an unauditable memory in an engineering loop means you don't know if the agent is working from current facts or stale ones. 👀

Dude, this is real alpha, I respect this guy

Yeah true. The guy is a real chad. Very useful watch in my opinion for those who are building agentic systems

nice one.

Thanks Alexa !

@0xCodez honestly, thought it'd take way longer than 19 mins. gonna try this loop next project.

When agentic loops execute long-running tool calls over episodic memory tiers, trapping your state in ephemeral wrappers is why production systems break! True autonomy requires an infrastructure that treats agent trace and memory not as background logs, but as a core, queryable data layer.

This post contains real value

Thanks bro. Hope people will use at lest 10% what’s explained in this video

biggest surprise i had building loops was how much the eval quality matters more than the harness itself — bad evals just accelerate the wrong behaviors

This video was originally created by @ShenSeanChen and was re-uploaded without his permission. Re-uploading someone else’s original work without permission is copyright infringement, not “sharing”. Please delete this post and quote the original post instead.

plll mi mi mi mmm m mi Mi i nul Ok lo lllll poi lo puoi lavare

The evaluate step is where most agent loops quietly break: it assumes the trace is readable enough to judge while the agent is still running. If the output keeps shifting or lands as an opaque block, evaluation becomes a post-mortem and the loop never actually self-improves. It is the same readability-under-streaming bar we work on with Lanson Live, keeping live captions stable enough to follow as speech happens instead of just fast.

building robust evals is the hardest part of agents

awesome

trace → evaluate → diagnose → fix → ship is the whole product most teams stop at spawn agent and never ship the eval layer

The focus on evals is refreshing. A lot of agent projects spend months adding features before they build a way to measure whether anything is improving.

That loop is the entire harness. Trace, evaluate, diagnose, fix, ship - every step codified as a skill so the next run inherits it. How SKILL.md fires without prompting:

This is worth looking into, thanks for sharing

Running this exact loop across 21 specialized agents at a PE firm. The eval step is underrated — knowing when an agent's output is portfolio-grade vs. plausible-but-wrong is where most firms stall. Memory across sessions is the next unlock.

trace, evaluate, diagnose, fix, ship, repeat that's just how senior engineers debug, finally written down for agents

"level‑up-ing with style" is doing a lot of heavy lifting in a sentence about observability pipelines lol

the trace and ship steps are new to me.

Trace → evaluate → diagnose → fix → ship → repeat i love this formula

Memory + evals are where most agent stacks fall apart. Good to see them front and center here.

the eval loop is the part most people skip. they build the agent, ship it, then wonder why it degrades over a week. trace → eval → fix is literally the difference between a demo and a product. been running this exact cycle daily with my agents and it compounds fast

every framework tutorial skips the harness and goes straight to demo which is why 90% fail in prod

Nice breakdown. Love the pragmatic loop—trace, evaluate, fix, ship. Agentic loops with memory and evals sound like solid paths to continuous improvement. Congrats to the team for framing it so clearly.

self-improve is lowkey a stretch, that loop is only as good as its evals, unless they catch regressions it just repeats the same bugs faster 🤔

This is where many agent demos become real systems. Logs are useful, but the production leap is structured traces: intent, tools, state changes, approvals, failures, and recovery paths.

Trace, diagnose, fix, repeat is basically just good engineering applied to agents.

Codex has a slightly different take on loop engineering. I notice building a harness