Brendon Brown

Fractional CTO, digital strategist working with private brands, religious institutions, and mid-market businesses.

[Brendon Brown @brendonwbrown]

And this is only consideration towards supported enterprise and consumer infrastructure.

[Brendon Brown @brendonwbrown]

On the Microslop outage: AI gives indeterminate results which require multiple attempts, testing cycles, before it becomes usable. With the volume of legacy code and its history at Microsoft, only a heavy model with a massive context window could consider full potential impact.

[Brendon Brown @brendonwbrown]

@mattpocockuk "I bet you say that to all your other tools."

[Brendon Brown @brendonwbrown]

This is actually an @NousResearch Hermes appreciation post... in spite of appearances:

https://t.co/j5HqR3Kai6

[Brendon Brown @brendonwbrown]

Judge your bots by outcomes, not by how much they yap with each other.

There.

I said it.

>>>

[Brendon Brown @brendonwbrown]

My standard is a simple one: would the agent's work be measurably better if this role had every access, tool, and piece of information in reach? If the answer is no, then I pay the bot tax, accept the friction, lose some fidelity, and isolate the roles.

[Brendon Brown @brendonwbrown]

Judge your bots by outcomes, not by how much they yap with each other.

There.

I said it.

>>>

[Brendon Brown @brendonwbrown]

Risk exposure drops sharply when individual agents' access is scoped deliberately to their task list. You would never hand a new hire the CEO's credentials simply because it would be more efficient.

[Brendon Brown @brendonwbrown]

If you are building internal repeatable processes or systems, or building out features for other businesses, let me know and I will send you the checklist I tag onto the end of every review.

[Brendon Brown @brendonwbrown]

Derisk your agentic output by designating a swappable environment file to hold the personal and sensitive values. After this, hand your output to a detached reviewer with a four-line checklist that names what a leak looks like. If the reviewer has a /review prompt, use it.

[Brendon Brown @brendonwbrown]

I've been working with my agent to dump my hard-earned experiential processes into skill that replace some of the mundane things I have to do. Brain dump, plan mode, review plan, build, test. Thinks look good? Production.

[Brendon Brown @brendonwbrown]

Current issue: my agentic sessions are now highly personalized, and the code, skills and shareable content carry that personality. The facts that make my agent smart are the same facts that do not belong in a process meant to run on a stranger's machine.

[Brendon Brown @brendonwbrown]

Memory is a major reason my agentic AI work is fast. Sessions already know which model breaks below a certain quant, which utility/machine is off limits, current color, font preferences for a client; often the LLM infers what I want as soon as I finish sending a prompt.

[Brendon Brown @brendonwbrown]

Do you read your code before shipping? How about your Agent skills? Do you read them in their entirety? If not, which is many of us now, you have to make sure you are using review tools to their full, or you are headed towards a privacy disaster.

[Brendon Brown @brendonwbrown]

@minchoi @gregisenberg They are cooking, and we are so cooked.

[Brendon Brown @brendonwbrown]

@NetworkChuck