active
Capability vs. Control
I work with always-on agents like Grok Bot and Meta’s Muse and watch the control surface: shared credentials, standing approvals, and whether removing an agent actually revokes its access.
Independent researcher · Agentic systems · Cybersecurity
I build multi-agent AI systems,
then I try to break them:
coordination,
security,
and control.
exAgentica is my personal agentic AI lab.
I build experiments and write about what I learn. After three decades in Internet infrastructure, I keep coming back to the same questions: what can fail, how far can it spread, and who can stop it?
Read the Field Notes01 / The field notebook
Field Notes
What I’m learning about agents, security, and the infrastructure underneath.
A Bot Is Not a Bot Is Not a Bot
The word "bot" now covers a ticket scalper, a search crawler, a page fetched because someone asked, and an agent using your login. Lumping them together scrambles the traffic numbers and the blocking decisions. These are the questions that matter more than the word.
A Password Is Not Permission
Personal AI agents are being sold on convenience, down to a photo of your credit card, while the biggest banks lock them out of moving money. There are better ways to give an agent limited spending power, and I want to hear where you'd draw the line.
The Software Purchase That Never Happens
My AI subscriptions are billed to me as a consumer, but most of what they do is work organizations once bought software for. When an agent does the job, the product never gets bought, and nothing records the decision.
02 / Experiments
Active lab work
The questions I’m spending time on. Active work and research notes.
active
Multi-Agent Orchestration
I test how agents divide work, share memory, and recover when a handoff fails. When they disagree, I watch how they decide, spot circular work, and know when to stop. Coordination is the system, not the glue.
active
Agentic Coding
I use coding agents to write and review software. I look for the point where a passing test hides a decision that still needs human review.
active
Local LLM Deployment
I run models locally with Ollama and LM Studio, testing which tasks can stay on my hardware and what I give up in speed and capability.
active
AI-Driven Security
I study agents on both sides of a security boundary. Tool permissions, untrusted context, and response authority are part of the same problem.
notes
Post-Quantum & Zero Trust
I examine machine identity and cryptographic dependencies. Which permissions should an agent inherit, and which should it have to earn?
What I’m working with
- Claude Code
- Codex
- AgentsView
- Grok Bot
- Meta Muse
- Cursor
- Antigravity
- Ollama
- LM Studio
- TypeScript
- Python
- Cloudflare Workers
- Astro
- LangGraph
- CrewAI
- MCP
- Docker
- Git
- Linux
I approach multi-agent AI as a distributed systems problem. Tools, memory, and permissions make each agent part of something larger. A local mistake can become a system failure.
I co-founded one of the first ISPs in Texas, helped launch OpenStack and the Open Compute Project at Rackspace, and served as Field CTO at Cloudflare. That experience shapes what I test here.
Full bio at johnengates.com04 / Contact
Let’s compare notes.
If you’re building agent systems or working through AI and cyber decisions on a board, I’d be interested in a conversation.