Building agent infra for agents fighting cybercrime
Joined October 2008
- Tweets87
- Following472
- Followers143
- Likes594
Aaron Rolett retweeted
i feel like swe interviews could probably collapse into just 2 technical rounds:
• system design (how well can you articulate your ideas, do you actually know how to engineer something)
• onsite project to build a real thing with agents (can you translate your intent into high quality outcomes)
everything else we used to ask is no longer necessary
🗞️ if your team is too busy doing their 'normal job' to experiment with AI, you're preparing them to be replaced
ghuntley.com/replaced/
In this post, I guess this's the last time I'll say it. AI use is no longer optional if you wish to be employed. What follows is advice for current employees and hiring managers on how to hire in 2026 to source AI-first candidates.
Shifting left is the superpower of using AI tools in 2026.
If the finished product isn’t on point, throw it away and add more constraints to your concept/tech plan
Real anti pattern these days to watch agents code.
One of the questions I ask myself frequently:
Where am I being a subagent for Codex, and what do I need to build to get out of the way?
Building software to get yourself out of the loop is an extremely high-leverage activity.
You can’t defend against attackers moving at agent speed with human-speed response. Exciting use of behavioral modeling.
Last week, @Abnormal announced AI Cloud Security, and I wanted to share our internal prototype about how we’re using Abnormal’s core behavioral AI detection and response technology to stop the next generation of AI attacks.
AI Cloud Security is built on the same core architecture and technology that we use for our email and identity security products that protect ~30% of the Fortune 500 today:
1) Map all the human, non-human, and cloud identities
2) Build behavioral models to understand baseline / expected behavior
3) Detect behavioral anomalies
4) Respond autonomously at machine speed
We’ve spent the last several years building the core technologies, and I’m excited to start sharing our new AI security products as we build them with this audience. These products will be available to customers this year, but in the meantime, we are working with select design partners to ensure their success in the largest, most sophisticated enterprise environments.
We are learning and building every day, and I’m excited to share more as we go!
Migrated to a new laptop recently and realized the most prized data on my laptop these days is Claude/codex session logs and my agent plan specs.
Customer-obsession for platform teams at an AI-pilled company
Old way: engineer → docs / self-service → hits a bug → Slack the platform team
New way: engineer hits an issue → loads the platform skill into claude/codex → keeps iterating
The agent debugs alongside them, answers questions, opens PRs, or escalates with the context the platform team needs.
Gives a better customer UX and a foundation for a platform software-factory loop: support interactions turn into fixes, better tooling, and the platform backlog.
Aaron Rolett retweeted
we've built so many ways to kickoff our internal agents and yet slack just dominates for all tasks
like even for the (email) folks who dont even use slack that much
builders.abnormal.ai/p/nora-…
Agentic threat modeling: some agents get compromised. Others are malicious from prompt one.
Two attack vectors with very different risk profiles:
Trusted agents hijacked via prompt injection — benign agents manipulated into taking adversarial actions.
Attacker-controlled agents inside the environment — adversarial by design, able to recon, adapt, and move at agent speed.
One of the most valuable skills in a AI-pilled company is turning customer needs into engineering problems.
Now that implementation is cheap, judgement on what problems to solve and how to deliver it in a rapid feedback loop is more valuable than ever
frontier models are great at ASCII art. My favorite way to consume information for architecture, software factory plans and learning new concepts is to ask for before & after ascii art.
To get out of the coding loop:
1. Define what evidence makes a PR mergeable
2. Align on the plan
3. Give the agent tools to execute + produce that evidence
4. Review the evidence — don’t simulate the program in your head
For distributed systems, #3 is the hard part.
This! I find myself only involved in the dev loop at two points: aligning the plan and reviewing the testing.
Once you do this the loop collapses and the job becomes figuring out:
1. Is this the right architecture?
2. Does the testing convince me it works and avoids regressions?
