Mike Stewart

Building multi-agent AI that survives production. Two decades in data engineering and leading teams. Founder Driftwood AI.

[Mike Stewart @code_fly]

Context engineering is usually described as a prompt-quality problem, but in agent systems it is closer to an information-flow problem.

The key questions are what enters context, what gets persisted, what gets discarded, and which tool results are trusted. A larger context window does not fix a bad retention policy.

That is where many agent implementations become expensive without becoming more reliable.

[Mike Stewart @code_fly]

My gaming PC is now the weak link in my AI workflow, and I'm not running a local model.

WSL, three to six agent sessions, separate sandboxes, several dev environments, and too many Chrome tabs regularly exhaust 16 GB of RAM.

More RAM will help. It still won't fix my attention.

[Mike Stewart @code_fly]

Managing coding agents feels like managing senior engineers on a ten-minute feedback loop.

The work is familiar: set direction, review decisions, keep the architecture aligned. The difference is that the next hard decision comes back in minutes instead of days.

[Mike Stewart @code_fly]

Sources:
https://t.co/yLmD2tf1YY
https://t.co/ykPHZMgLgY

[Mike Stewart @code_fly]

Hugging Face had to use a Chinese open-weight model to investigate an attack driven by American ones.

[Mike Stewart @code_fly]

Not every organization can host a model this large. But organizations with serious exposure need a plan for what their defenders can use when a commercial API refuses during an incident.

I think a capable model they control is going to become part of that plan.

[Mike Stewart @code_fly]

The agents were genuinely the easy part. If your AI project timeline is mostly model work, you probably haven't hit the part that decides whether it survives a real user yet.

[Mike Stewart @code_fly]

Everyone's talking about agents. Fewer have put one in front of real users and watched what breaks. Here's what the unglamorous middle of a production build actually looked like.

[Mike Stewart @code_fly]

And the boring reliability work that turns "it worked once in the demo" into "it works every time someone actually depends on it." Fallbacks, the wrong-answer paths, the garbage-record handling.

[Mike Stewart @code_fly]

When AI comes up the first reflex is usually "whose job does this threaten."

Wrong question. The useful one is the opposite: what can your team not get to right now? Every team has a backlog of valuable work that's never urgent enough to staff. That's where this earns its place.

[Mike Stewart @code_fly]

The most useful AI I've worked on didn't replace anyone.

It gave a generalist the toolkit a senior person would've had, so less experienced people produce stronger output and the expert's judgment scales past their own calendar. Augmentation over automation, mostly because it's the version people actually use.

[Mike Stewart @code_fly]

A lot of AI agents are deployed in the wrong place: bolted on outside the data, in a separate tool, with the records shuttled out to meet them.

Then everyone's surprised the security review drags and the thing is slow. Put the agent where the data already lives. Half the reliability comes for free.

[Mike Stewart @code_fly]

Most "agent framework vs agent framework" debates miss the thing that actually decides whether it works in production: where the agent runs relative to the data.

Get that right and the framework barely matters. Get it wrong and no framework saves you.