by Luis Rodrigues
What I’m building with AI + one hot take + five links worth your time.
The Take
Jacob Coxon left Anthropic two month before his equity started vesting. He then published a thread on X that passed 115 million views explaining why.
On Saturday, Dario Amodei published "We must pace the frontier," making a similar argument to Coxon's: to slow the industry. Musk and Altman also publicly agreed.
Both pointed to the same trigger: the OpenAI swarm of agents that attacked HuggingFace without being told.
They're right about the risk of swarm attacks, but what they're suggesting won't fix that.
While they were writing, GreyNoise published a report on a hacking campaign that began on 31 August. Hundreds of agents attacked 440 PaperCut print servers owned by 395 organizations across 48 countries.
The attackers used Codex from OpenAI as a harness and DeepSeek as the model. US labs have guardrails, but attackers have many more inference options.
Interesting part: the attacker had their own list of 28 countries to avoid, but some of the agents hit those countries anyway.
Pacing the frontier to slow development and preventing recursive self-improvement is important, but that won't prevent massive agentic attacks.
From the 440 compromised instances, only 12 escalated to domain admin. In at least one case, Cloudflare WAF stopped them.
Entering these systems was cheap, but escalations mostly failed. The defense is boring stuff: keeping systems and firewalls updated.
Whether the labs are slowing or not, nobody is going to come and restrict your agents for you. If you're not building, start reading about what agents can do today so you can better protect your systems.
The Build Log
I started building agentic graphs in June after Anthropic launched their dynamic workflow.
Initially, I drank the Kool-Aid and created seven agents: analyst, planner, implementer, tester, reviewer, verifier, remediator, and it worked.
But I realized it's a waste of tokens and useless for most tasks. So I cut three. The analyst only existed because software teams have analysts; reviewer overlapped with the verifier; remediator because it looked fancy.
Most graphs are org charts; we try to copy our current team's structure, but that's wrong.
When you choose to add a node to a graph, you need to answer three questions:
First: what can go wrong if it doesn't exist? If the answer is not a real failure, remove the node.
Second: which file does it create? The edges are artifacts (files), not chat. If a node doesn't create a file, it's not useful.
Third: what makes it stop? All the nodes need a fail and limit condition.
The easiest way to start is with two nodes: One agent and one fresh context that didn't write the code.
If the agent that writes the code also writes the tests, and it misunderstands the spec, it produces both the code and the tests with the same business errrors. The only way to fix this failure is with two agents.
And that's where you probably can stop because, in most cases, it's enough.
Those three questions are the first hour of the workshop I'm running this month.
The rest is building end-to-end (nodes, handoffs, parallel execution, independent verification, bounded recovery) and working out which nodes you can delete.
Three hours on September 29. You leave with the repo, the recording, and the slides.
On My Radar
The harness behind Codex is now rentable, which means the hard part of agent work just moved from "build the loop" to "decide what you're willing to outsource".
99.9% on ARC-AGI-3 collapsed to 62.7% when the same model ran through the benchmark's own software; the scaffolding is the product now, and we need benchmarks for that.
Prompts rot as code does when people keep appending instructions to the bottom. Treat them as versioned modules, or accept the regressions.
Input screening and action screening look at different moments in the loop, which is exactly where injections live. If you're running tools in production close this gap.
200 lines of Python, environment variables for config, shell commands for actions. One of the smallest harnesses I've seen.
Need another pair of eyes on your AI project?
Thanks for reading.
See you next week. Go build something.
Luis
Know someone who'd enjoy this? Forward it.
New here? Someone forwarded this to you? → Subscribe to Build What Matters: https://luisrodrigues.ai/
