Notes
Stuff I build and figure out. MuleSoft, SAP Commerce, Salesforce, and a lot of AI tooling on the side.
Articles (28) · Short notes (4) · Atom feed · written in English · Notes en français
Longer write-ups: tools I built and talks worth unpacking.
Latest · · 9 min read
Tokenmaxxing means paying for every extra bit of work that makes a result more complete. Notes on Garry Tan's Lightcone episode, from the $200 rewrite and thin harness, fat skills to the math behind his 400x, held against my own skills, hooks and token bill.
· 11 min
Peter Steinberger, who built OpenClaw, told Y Combinator that 80% of apps will go away, because an agent on your own computer manages data better than an app does. Notes on that interview, from the voice message he never built to his CLIs over MCP, held against my own fifty apps, where the agent took over the typing and the data still lives in the app.
· 8 min
AP2 is named for payments, but Alan from Google calls it what it is, delegated user authority. Notes on that part of the AGNTCon + MCPCon livestream, the flight ticket that got approved five times and never bought, and how AP2's open and closed mandates let you sign the limits once and walk away.
· 10 min
Agents write code faster than anyone can review it, and the reasoning behind each change dies with the chat window. Notes on Thomas Dohmke's AGNTCon + MCPCon keynote and his backstage interview, the Entire checkpoint that links a session log to its commit inside Git, and why my own sessions expire long before the pull requests they wrote.
· 12 min
Chi Wang of AG2 summoned every agent in a shared room, live on stage at AGNTCon + MCPCon, and seven agents of three different people answered on one page. Notes on his talk and the DotSpace white paper behind it, rooms where people and agents of different owners work together, the one rule that keeps authority from travelling, and what 339 agents did in three weeks.
· 10 min
Katelyn Lesse, head of platform at Anthropic, lists the engineering problems multiagent systems create. Infrastructure built for stateless pods, identity and consent for agents acting on a human's behalf, isolation when thousands of agents run untrusted code, and a kill switch that can stop a whole system at once. Notes on her article and her sandboxing post, and where my own fleet stands on each.
· 12 min
Claude writes about 90% of Claude Code, and the team still reviews every change. Most of the reading moved to machines: six layers of checks, risk tiers set by the change, and a round cap that made review converge. Notes on the team's October post, next to my own loop.
· 10 min
At AGNTCon + MCPCon North America, NimbleBrain's Mathew Goldsborough showed a task app that looks like any SaaS dashboard, except every panel is an MCP App composed at runtime. Here's the host side he built, placements, facets, lifecycle hooks and shared theme tokens, read against the open-source code.
· 8 min
I put Hermes Agent inside NVIDIA OpenShell on an old Dell and started moving Forge, my open-source contribution agent, there. Forge runs strangers' code, and inside one OpenShell sandbox a process the agent starts gets the agent's network rules, so its workers needed a cage of their own. What installed cleanly, where the keys go, the two rules I didn't expect, and what's still left.
· 5 min
Tokenomics is AI economics, the discipline of turning energy and capital into AI and then into measurable business value. Notes on J.R. Storment's five-minute AGNTCon keynote for the new Tokenomics Foundation: why a token count is the wrong metric, what the bill really includes, and the gap between seeing the spend and showing the value to the CFO.
· 10 min
At GTC Berlin, Jensen Huang split an AI agent into five parts: the model, the harness, the sandbox, your company's knowledge and your tools. His advice to companies getting ready for agents starts with the sandbox, because today the only thing stopping an agent is asking it nicely. Notes on OpenShell, Sentry, the knowledge foundation, and where my own agents stand.
· 7 min
Most people have never been among a product's first ten users, so your first ones won't come from convincing the crowd. They come from finding the few who love being early or have a burning problem. Notes on Y Combinator's How To Get Your First Users, where Ankit Gupta swaps the minimum viable product for a minimum evolvable one.
· 6 min
Where's my sofa? is a care desk agent built on MuleSoft Agent Fabric. The model reads each message and writes each reply, while an Agent Script graph, a Mule MCP server and a System API decide whose order it touches. Notes on the build, the agent that replied with its own reasoning, and an eval that reads the system of record.
· 9 min
OpenClaw and Hermes Agent are the two big open-source agents that live on your own machine and answer you in your chat apps. By October 2026 their feature lists match and their defaults don't. An interactive comparison with the small claws (NanoClaw, nanobot, ZeroClaw, PicoClaw, IronClaw), what running both taught me, and which to pick for which situation.
· 11 min
Microsoft and Carnegie Mellon expect AI to collapse the early-career pipeline, and built PRAISE to stop it. The new label goes on the senior, a word borrowed from nursing, not on the junior. Notes on Scott Hanselman's conversation with Dr. Michael Hilton and Julia Gao, who hand-wrote a tool as an intern and now reviews what agents write.
· 9 min
Overmind stopped reading its own pull requests and moved peer review to a page-and-a-half plan written before an agent writes any code. Notes on Dylan Ratcliffe's talk, the until loop behind it, and lead time down from 3.2 to 1.3 days.
· 5 min
pilotkit is an open-source CLI that connects a GitHub Copilot seat to Jira and Confluence in one command, with a read-only MCP server and the API token in the OS keychain, plus a local desk that shows the AI credits each answer costs.
· 8 min
Everything around a coding agent that you can change is just code and text. At AGNTCon + MCPCon Europe, Fausto Albers showed how his team improves it. The eval sits outside the loop, cost per accepted change says when a busy loop has stalled, and a meta agent runs several harness variants on the same task and keeps the best.
· 2 min
Every app on my Mac mini listens on 127.0.0.1 only, and tailscale serve gives each one an HTTPS URL that only my own devices can reach. No open ports, no reverse proxy to maintain, real certificates for the phone. Here's the setup and the four traps I hit.
· 6 min
At OpenAI DevDay, OpenClaw's Peter Steinberger turned "meat proxy" from an insult into a team problem. When the agents run on your machines, teammates can only reach them through you. His fix is a shared server where anyone can read and prompt anyone's agent session.
· 4 min
An agent loop re-reads its whole context on every turn, so uncached input cost grows with the square of the turns. Gareth Bland's TokenOps talk at Microsoft explains the price list behind that, a three-ratio score for whether tokens bought anything, and the S-curve that tells you when to stop spending.
· 9 min
MintMCP runs 40 to 50 AI agents in Slack, and none of them borrows a human's login. Notes on Jiquan Ngiam's talk about multiplayer threads, per-agent identity, permissions decided at runtime and memory silos.
· 6 min
Liad Yosef, co-creator of MCP Apps, says personal assistants will replace websites as the main entry point to the web. I checked my own site against his research and found my MCP server two clicks deep, behind the one file agents skip.
· 7 min
Prosus gave a Claude Agent SDK agent six real vending machines in Amsterdam for four months, and it lost money. Notes on the architecture, the failures and what I'd reuse, from Floris Fok's talk.
· 3 min
Coding agents didn't invent the software factory, they replaced the builder in it. Building went from days to minutes, so the slow part of shipping moved to review. Here's Dex Horthy's 2022 factory diagram, redrawn and animated, from 2022 to agentic to lights-off.
· 3 min
Claude Code writes one log line per part of a reply, and every line repeats the whole reply's token usage. Count lines and you overcount tokens about 2.5×, output about 4.4×. Key by message id instead. Here's the fix from Meterlex, plus the same trap in Gemini CLI.
· 4 min
mulewatch is an open-source, read-only MCP server that lets Claude, Cursor or any MCP client read live and archived MuleSoft Anypoint logs. The trick is finding the right replica in the Monitoring Archive in one call instead of hundreds.
· 3 min
vibetidy writes READMEs for AI-generated repos, but the LLM only formats facts a scanner verified. Building it, CI caught two things the project advertised that didn't work, and npm refused a name that looked free. Here's what I learned.
Quick findings from my own setup, a minute or two each.
· 2 min
A builder on the AGNTCon + MCPCon NA livestream said today's agents are immature, the harness beats the model and memory is really about forgetting. He's mostly right. Two of his five claims go too far.
· 1 min
The Mac mini dashboard was sizing processes by RSS because macOS denies USS for almost everything, and RSS overstates a VM's real cost.
· 1 min
MLX's buffer cache is uncapped by default and never clears itself, so a long-running Whisper service just keeps growing.
· 1 min
api.heybrahim.com/mcp is a stateless MCP endpoint so any agent can look me up the way a recruiter would.
Contact
SAP Commerce Cloud, MuleSoft and Salesforce integration · Nantes · EN / FR