itsbrill.ai

How I work

My agents do most of my knowledge work, across my job at an AI company and my own projects. This is what that looks like, and the systems I built to make it work.

What it adds up to

My agents do most of my knowledge work now, across 8 workstreams and 22 active projects in different departments. Most days I’m the meat proxy, the person in the room for work my agents have done. I’d call it the work of about ten people.

  • 22active projects across 8 workstreams
  • 3,602agent sessions remembered since March
  • 3,651sources filed, from emails and Slack to meeting transcripts and documents
  • 111meetings turned into notes, decisions and actions
  • 81disagreements between sources caught
  • 33shared skills every agent works from

Counts from my own system on 30 October 2026.

The hard truth

Agents can now do most of the knowledge work one person used to be hired for. Mine already do most of mine, and I don’t think a job built around a single function survives that.

I think companies will keep cutting roles, and the people who stay will be expected to run a fleet of agents across several areas at once, the way I run mine across eight.

The companies doing this well aren’t scaring their people into it. They’re training them, and giving them agents worth using.

The good news is it’s still early. Most people I talk to haven’t started working this way yet, so there’s time to get ahead of it, but that time is running out.

That’s why people message me. They want to be ready, and to start working with agents the way I do.

Work with me

What a fleet of agents can do

Each of these is a fleet of agents splitting one job between them, with how it’s gone in my own work underneath, names taken out.

Presentations

A QBR deck in under two minutes

It’s 4pm and the QBR is tomorrow. Instead of spending hours lining up brand assets and copy, your fleet builds the deck from your numbers and your template, so your evening goes on preparing what you’ll actually say.

Don’t worry designers, you’re not being replaced. You now have a teammate you can delegate picture-perfect designs to.

From my work

Customer, partner and internal decks for my day job, built the same way.

Watch the fleet build it

11 slides in 1 minute 42 seconds, every figure checked

Overnight

Wake up to work done

Before you log off, you hand your agents the light work. They get through it overnight, and when you open your laptop it’s done, with a one-page brief on what’s overdue, who’s waiting on you and the three things worth doing first.

Finally, a colleague who’s happy to work the night shift.

From my work

Builds and research run overnight, and every day starts from a brief across 22 projects.

The fleet

  1. Researcherworks through tomorrow’s prospects
  2. Writerdrafts the replies you’ve put off
  3. Builderfinishes the job you started
  4. Chief of staffwrites your morning brief

Done before you wake up

Security

A 300-question security questionnaire, done by Friday

A customer sends a 300-question security questionnaire on a Friday afternoon. Your fleet answers everything your policies already cover, quotes the source for each answer and leaves you only the ones that need a real decision.

Compliance team, you can have your Friday back.

From my work

A 129-item security audit, with the evidence drafted and every owner tracked.

The fleet

  1. Coordinatorsplits up the 300 questions
  2. Answer agentsanswer from your policies, quoting each source
  3. Reviewerchecks every answer against its source

A dozen left for you to decide

Launches

A launch kit in a day

You’ve got a launch next week and nothing ready. You brief one agent, it briefs the rest, and by the end of the day you’ve got the research, the announcement, the FAQ, the graphics and a landing page to look over.

Middle management, meet your new direct reports.

From my work

This website. One agent designed it, another built it, a third coordinated, and the contact form is answered by my chief-of-staff bot, Frank.

The fleet

  1. Chief of staffbriefs the team
  2. Researchersizes up the competitors
  3. Writerdrafts the announcement and FAQ
  4. Designermakes the graphics
  5. Builderputs up the landing page
  6. Reviewerchecks the lot

Launch kit ready by the end of the day

Accuracy

The contradiction caught before your customer sees it

Your sales deck says one thing and the product docs say another. Before anyone quotes it to a customer, your fleet spots it, quotes both and asks the right person which one is true.

The “final_v3_ACTUAL” file can finally rest.

From my work

81 caught so far, like a partner deck and the product docs giving different driver versions.

The fleet

  1. Watcherreads every new doc, deck and message
  2. Checkercompares it with what you already know
  3. Writerdrafts the question to whoever owns it

Fixed before a customer sees it

Training

Train 200 people by Friday

A new tool goes live on Monday and 200 people need to know how to use it. Your fleet writes the guide, turns it into a cheat sheet, slides and a quiz, then answers everyone’s questions from it afterwards.

Your inbox survives go-live week.

From my work

AI enablement for a whole company, and the free guides on this site.

The fleet

  1. Writerwrites the guide
  2. Designerturns it into a cheat sheet and slides
  3. Buildermakes the quiz
  4. Assistantanswers questions from the guide afterwards

200 people ready by Friday

Product

From customer call to working prototype

Every customer call is full of ideas. Your fleet pulls them out, builds a working proof of concept from the best one and hands it to your product team in the format they already use, the spec, the user stories and the demo. It works best in a flat team, where a good idea can go straight to the people who build.

Product managers, the backlog just got a lot more interesting.

From my work

A team knowledge base I built as a proof of concept, now handed to product and engineering to build on.

The fleet

  1. Listenerpulls ideas and pain points from customer calls
  2. Researcherchecks what already exists
  3. Builderturns the best idea into a working proof of concept
  4. Writerpackages it the way your product team works

Ready for product to build on

And plenty more

  • Meeting notes and follow-ups before you’ve made a coffee
  • Emails in your own voice
  • A new starter who doesn’t need to ask
  • It remembers why you decided that
  • Competitor research before lunch

Why my agents don’t make things up

Every fleet I run sits on two systems I built. The first is agent-si, a brain that only holds facts quoted word for word from a source, and every agent checks it before it answers. If it isn’t in the brain, the agent says so instead of guessing. For me, it’s essentially eliminated hallucinations.

When two sources disagree, it flags both rather than picking one, and it’s caught 81 so far. In my test, plain notes repeated a planted misleading value in 12 of 72 answers. agent-si never did.

The second is agent-memory, which remembers every session, decision and lesson across all my agents. agent-si knows what’s true and agent-memory knows what happened, so any agent can pick up where another left off without me explaining it again. On a public memory test it found the right past conversation in its top 5 for 95.8% of 500 questions, with no AI model.

That’s also what takes the load out of switching. Every project’s state lives there, so when I move between them, an agent briefs me on where things stand and we carry on.

See all the test results

Under the hood

Four systems sit under every fleet, and each one has its own story below.

skills
Shared instructions for how a task gets done, read by every harness before it starts.
agent-memory
One local search index of past sessions and notes, so an agent can find what we already worked out.
agent-si
Facts only, each quoted word for word from a source.
agent-msg
A small message server on my Mac, so Claude Code can hand work to my Grok Bot team.

Every agent works from the same system

For your teamEvery AI tool your team uses works from the same instructions and the same knowledge, so switching tools doesn’t mean starting over.

1The problem

Each harness starts with its own memory and its own habits, so an agent in one has no idea what I worked out with an agent in another.

2What I built

Every harness shares 3 parts. The history comes from agent-memory, the facts from agent-si and the skills from my shared skills repo.

3How it stays right

An agent only writes a step up as a skill or a note once it has actually worked.

4What surprised me

Helper agents don’t go looking for skills on their own. I had design helpers build mockups and none of them loaded the design skill, because nobody told them to, so everything came back generic and I rejected it. Now every task I hand a helper names the skills it has to load.

The problem

Agents find what they already learned

For your teamYour team’s AI tools remember what’s already been worked out, so nobody solves the same problem twice.

1The problem

My agents kept re-learning the same issues, even though the answers were already in my notes, because nothing was searching them.

2What I built

Sessions and notes from every harness go into one local search index on my Mac.

3How it stays right

Lessons only get saved at a checked point, like a reviewer signing off or me giving an instruction, and agent-memory refuses to save one without evidence attached.

4What surprised me

Some searches kept surfacing the same useless matches, so I added an exclude list. Lessons were drifting into status updates, so I tightened the rule to lessons only. The biggest surprise was how far plain keyword search goes. On LongMemEval, a widely used public test for AI memory, it found the right past conversation in the top 5 for 95.8% of 500 questions, with no AI model.

The problem

How agent-memory brings back what an agent already learned.

agent-memorySearch
Design helpers came back generic.

Lesson

Name the skills in every helper’s task.

Agents answer from quoted sources

For your teamAI that answers from your own documents, quotes its source and tells you when two documents disagree.

1The problem

I got sick of AI models making up answers.

2What I built

agent-si is a knowledge hub that only holds facts quoted word for word from a source. When two sources disagree, it keeps both and flags the dispute.

3How it stays right

If an agent wants to overrule a fact another agent saved, the change waits for me.

4What surprised me

In my test, the notes repeated a planted misleading value in 12 of 72 answers and agent-si never did. My first full run also failed some checks because my grader was marking right answers wrong, so I fixed the grader before I trusted any of the numbers.

The problem

How agent-si keeps a quoted fact and flags a dispute.

agent-siQuoted fact
The Pro plan includes 10 seats.

Dispute

An older sales deck says 5 seats.

Both values are kept and flagged.

Claude Code hands work to my Grok Bot team

For your teamAgents in different tools hand work to each other, so a job keeps moving without anyone copying it across.

1The problem

My Grok Bot bots already talk to each other, but Claude Code had no way to reach them, and I didn’t want to copy messages between windows.

2What I built

A small message server on my Mac and a command line tool let Claude Code send a message to Frank, who hands the work on to the right bot.

3How it stays right

A read-only page on my Mac lets me watch the conversations, and I step in through my own chat.

4What surprised me

Nothing wakes Claude Code when a bot messages it, so a message waits until a session checks in. I cover that with a check at the start of each session and a banner on my Mac after 3 minutes.

The problem

How Claude Code reaches the Grok Bot team.

Claude Code Frankagent-msg
Claude CodeThe pricing page has a typo in the plan name. Please pass it to the right bot.
FrankPassed to the writing bot, and I’ll tell you when it’s fixed.

Work with me

Let’s get you and your team working with agents the way I do.

Work with me
Enter to openEsc to close