Your AI cost report shouldn't wait a quarter for plumbing
You don't need a data platform to start tracking AI spend. Pull usage from each tool's API, apply simple rules, and build a baseline now.
Blog
Notes from the lab.
Posts by Jerzy Filatow, collected from LinkedIn and X and kept here, with the date each one first went out.
All posts
-
-
Grok Bot, dots and Duet: who does what
The real decision for leaders is what it may do without asking you first.
-
2027 is the year everyone starts talking to their agents. Out loud.
Three calls for 2027: voice becomes the way we work with agents, agents that use your computer go mainstream, and passive wearables capture the context. Plus the one company I would bet on.
-
Banks blocking AI agents was about as surprising as rain in Melbourne.
Banks are pushing back on AI agents that log in as you. The real gap is agent identity: a way for an agent to prove who it acts for and what it may do.
-
AI agents can now sign in, pay and work inside your apps for you. So the question moves from "can the AI do it?" to "what did we hand over, and where does a person still say yes?"
Cursor Vault lets agents hold your logins and pay for things inside your apps. What matters now is who holds the keys and where approval sits.
-
Write down how your team works once, and it now goes wherever your AI goes.
Agent skills moving to an open standard means a team can write its way of working down once and carry it across tools. The playbook is the asset.
-
You've scaled your agents. Great.
Agents now commit and merge around the clock, and git was built for humans. Why agent first source control is coming sooner than most teams expect.
-
A million lines of code (tests included). 1,238 pull requests. Seven reverts.
Cockroach Labs ran coding agents like a teaching hospital. The lesson: when agents write more than anyone can read, check that every step happened.
-
My house is full of connected devices. What it's missing is the layer between them.
Gemini agent aims to be the layer that carries context across work tools. Why that in between layer, with spend caps and admin controls, is the real product.
-
Most of the internet still expects a human with an inbox at the door.
Giving an agent its own email closes the loop on sign ups, codes and receipts. Why that plumbing lets agents take whole jobs, and why sends still need approval.
-
Getting an agent to read X just got easy. Teaching it what to ignore is still the job.
Reading X is now easy for an agent. The hard part is the filter: what counts, how much comes back, and the rule never to post.
-
Take the labelled parts diagram from a technical manual. Imagine pulling it apart with your finger.
GPT-6 can answer with diagrams and sliders instead of text. What that means for training, industrial design and how people learn.
-
Your model is not the part of your agent setup you actually own. Your playbook is.
A big open weight model is only an easy swap if your agents run on written skills and procedures. Swap the engine, keep the playbook.
-
AGENTS.md was step one.
Coding agents need a written workspace they consult and update, plus a rule for retiring what is stale. Consult, update, retire.
-
The agent that builds its own plugins still opens a skill file first.
Even an agent that writes its own plugins starts from a written skill. Self extension does not end harness work; someone still owns the code.
-
Grok Bot and OpenAI Dots look like they are pointing at the same thing.
Always on agents with their own computers just became a category. Why this is harness and operating system evolution more than a model story.
-
Your coding-agent bill is a harness tax, not a model tax.
A Berkeley study shows the harness around a coding model can move cost far more than results. What to measure before blaming the model.
-
Same model. Eight harnesses. Twenty-five tasks.
With the same model, success swung widely across harnesses. The harness is where teams win or lose.
-
While visiting my cousin, who happens to be a surgeon, we got onto AI.
A conversation with a surgeon about where AI helps in medicine and where trust and accountability still sit with people.
-
There’s a new way to do AI in town, and it’s not RLHF.
A newer training approach is moving past RLHF. What changes when models learn from outcomes rather than human ratings.
-
Michael Burry just called BS on the AI slowdown.
Michael Burry argues the AI slowdown story is wrong. The real limits are power, chips and labour, not demand.
-
Two weeks on Grok Bot. One giant agent lost. Specialists won.
Two weeks running a fleet on Grok Bot. One giant agent lost; a team of specialists with clear lanes won.
-
Armin Ronacher ran a weekend "software factory" on GPT-6 Astra.
A weekend software factory burned a billion tokens and produced nothing shippable. Why the harness needs a kill switch, a budget cap and readable diffs.
-
Time for the dark factory?
Physical dark factories have run for years. Are agentic software dark factories overdue, and what checks would make them safe?
-
Still context engineering bro? That's so mid 2026.
Prompt, context, compound, graph: the labels change fast. The useful idea is making the shape of the work explicit, and keeping the whole stack.
-
Humans were 'ere 2026
Some engineering teams now ban code comments for agents. Who is a comment for, and what should replace it?
-
Yesterday Google put me in a captcha loop. Five minutes of prove-you’re-human. Still declined.
Agents act from datacentres, and identity providers treat that as suspicious. Harness choice is now about lock in and blast radius.
-
Let's talk about Cursor
While the giants competed on intelligence, Cursor competed on integration. A look at how it got there and the bet others missed.
-
Anthropic released Claude Fable 5 yesterday. Their most powerful model, public for the first time.
Fable 5 is powerful and expensive, and it keeps your data for 30 days. When does thinking cost more than the output is worth?
-
For engineers, Harness Engineering is becoming an essential way of working.
Every codebase will end up with a harness. It brings consistency and faster onboarding, and it gets the codebase ready for agents.
-
Why It's Time to Get Your Hands Dirty
Senior leaders who stay on the sidelines with AI will fall behind. The people experimenting today will lead tomorrow.
-
Are you missing the AI wave?
A surfing analogy for AI adoption: read the conditions, paddle early, and commit when the wave comes.
-
Prompt Padawan - The Journey to Jedi
An early guide to writing prompts that work, from simple questions to structured instructions.
-
Taking the Red Pill: A Guide to Building the Matrix for Dummies (or Let's Just Settle with an LLM)
A plain guide to bringing a large language model into a business: set goals, pick a product, and mind compliance and security.