Dave Lee

Building MEGA — the optimization layer for production agents.

[Dave Lee @mega__dave]

The prompt is usually the first thing people fix.

But in real agent systems, the failure is often somewhere else: evidence selection, context flow, tool calls, routing, workflow structure, or evaluation.

That is why agent optimization cannot stop at the prompt.

[Dave Lee @mega__dave]

https://t.co/WaPIZZGaZw

[Dave Lee @mega__dave]

This feels right.

AI pilots don’t usually fail because the demo looked bad.
They stall because teams can’t measure quality, trust, cost, and risk well enough to put agents into real work.

Evals are the starting point.
The post-deployment loop is what turns those evals into continuous improvement.

[Garrett Lord @GarrettLord]

https://t.co/EBcWLBGkjz

[Dave Lee @mega__dave]

Models will keep changing. But a company’s workflows, evals, traces, and judgment shouldn’t reset with each new model.

That’s what needs to compound.

[Satya Nadella @satyanadella]

https://t.co/vLmiBKTtX3

[Dave Lee @mega__dave]

Is your agent actually driving business outcomes?
And do you have a loop that learns from real production data and improves itself to hit those outcomes better?

If this is something you’re running into with production agents, I’d love to connect.

[Dave Lee @mega__dave]

@marclou Thanks for hosting! It was great meeting everyone.

[Dave Lee @mega__dave]

@JunaidAckroyd Happy to connect!
Building Agent System Optimization.

[Dave Lee @mega__dave]

@mattlam_ @cursor_ai @FactoryAI @cognition @verrsane @verrsane Could you share plz?

[Dave Lee @mega__dave]

Interesting read. I agree that the real work starts after the trace.

Curious what felt different from other observability tools like LangSmith or Arize in actual use.

Also wondering whether fixing prompts and code is enough, or if the whole agent system needs to improve together.

[Dave Lee @mega__dave]

How are you validating your harness today?
If this is a problem you’re thinking about, let’s connect.

[Dave Lee @mega__dave]

Agent builders, quick question:

- Which model are you building with?
- Which model are you running in
production?

[Dave Lee @mega__dave]

@synhershko Adding autonomy is easy. Making it reliable is the hard part.

[Dave Lee @mega__dave]

A patch fixes today’s bug.
Optimization prevents tomorrow’s failure.

[Dave Lee @mega__dave]

More context on MEGA Optimus here.

We don’t assume the prompt is always the problem. The goal is to use evaluation to decide which system change should actually be kept.

https://t.co/i3RjdxmBuF

[Dave Lee @mega__dave]

The prompt is usually the first thing people fix.

But in real agent systems, the failure is often somewhere else: evidence selection, context flow, tool calls, routing, workflow structure, or evaluation.

That is why agent optimization cannot stop at the prompt.

[Dave Lee @mega__dave]

https://t.co/WaPIZZGaZw

[Dave Lee @mega__dave]

@iKunalmathur Makes a lot of sense.

Love the focus on real usage and actionable insights over complex dashboards.

[Dave Lee @mega__dave]

Building AI agents is getting easier.
Improving them is still hard.

Looking to connect with founders and builders working on reliability, evals, optimization, token efficiency, or AgentOps.

Curious what you’re building and what you’re struggling with.

Let’s connect, learn from each other, and build better agents together!

[Dave Lee @mega__dave]

@JustJerry121 MEGA does not start from a fixed layer. It uses traces and eval results to identify where the failure most likely entered the pipeline.

[Dave Lee @mega__dave]

https://t.co/WaPIZZGaZw

[Dave Lee @mega__dave]

How are you improving agent performance or reducing token usage?

Curious what's been working for people building agent systems.