An evaluation framework and learning environment where agents get classrooms, curriculum, and feedback.
22+ production PRs triaged
Hey — I'm Brandon. AI engineer, New York.
AI & agent
infrastructure
that ships.
I build MCP tools and agent evaluations, then operate the real-time systems around them — from production WebSockets to release-quality gates.
The work
✦ the pedagogical lens
If agents are students,
If agents are students,
someone has to build the school.
I came to agent infrastructure from the classroom: I taught music and web development to more than 100 students, from elementary school through college, across four NYC nonprofits. Now I build for agents the things that help students learn.
Classrooms
Safe, repeatable environments where agents can practice, fail, and retry without burning the building down.
Textbooks
Skills, docs, and specs that state one thing one way — the curriculum an agent can actually learn from.
Sleep
Context that compacts and resets, memory that consolidates overnight — rest is infrastructure, not downtime.
Teachers
Evals, feedback, and human judgment that set the quality bar — including triaging real production PRs.
✦ currently working on
What I'm building right now
A server-authoritative multiplayer music platform, with production WebSocket infrastructure and monitoring.
100+ concurrent players
Reusable agent skills that declare and resolve their tool dependencies.
MCP Skills Working Group implementation
An adversarial testing skill that finds security and correctness issues before production.
WDE Break prototypes before users do
Detailed project notes
4 projects
School Core
The school for agents: an evaluation framework and learning environment where agents get classrooms, curriculum, and feedback. Its eval system triaged 22+ production PRs through failure-pattern distillation — the quality bar that gates releases.
- agent evals
- trajectory scoring
- failure-pattern distillation
- curriculum
22+ production PRs triaged by the eval system
Sound Royale
A server-authoritative real-time multiplayer music battle platform — the live system where I operate infrastructure in production: WebSocket broadcast layers, CI/CD gates, Discord OAuth, health checks, and monitoring on Railway + Cloudflare Pages.
- WebSockets
- Django Channels
- React
- Playwright E2E
- PostgreSQL
100+ concurrent players at sub-50ms latency
School Core Skills
The textbook side of the school: reusable skills that declare the tooling they depend on and resolve it. I'm spearheading the reference implementation for tool-dependency resolution in the MCP Skills Working Group, so skills can carry their own syllabus.
- MCP skills
- tool dependencies
- spec work
- skill portability
WG reference implementation for tool-dependency resolution
Worst-Day-Ever
An adversarial stress-testing skill that tries to break prototypes before production does: security and correctness audits, edge-case abuse, severity-graded findings, and ELI5 explanations. It teaches agents (and their humans) where the weak joints are.
- adversarial testing
- state machines
- authZ
- audit reports
WDE break it on purpose, before users do
✦ contributing to
10 merged upstream PRs across the agent stack
Protocol specifications, runtimes, and agent automation — the plumbing that makes external agents trustworthy.
M
MCP Rust SDK
Protocol specs · SEP-2106 · SEP-2640
Relaxed the output schema type gate to accept arrays, primitives, and compositions while keeping inputs strict — 17 new tests, 424 passing, merged upstream after review.
PR 895 ↗O
OmniRoute
Agent runtime infrastructure
Resolved a 64× budget-multiplication defect that caused cascading service failures under concurrent agent load, and deployed a fair-scheduling budget that restored capacity. Also built Notion/Obsidian connectors with scope-enforced OAuth and closed an authz gap on 24 dynamic tools.
budget fix ↗O
Orca
Cross-platform agent automation
Eliminated a prompt-delivery race that left agent automations idle at startup — verified across five platforms and shells (POSIX, PowerShell, cmd, SSH, WSL) to restore reliable initialization.
race fix ↗
✦ writing
Notes from the agent classroom
I write about the pedagogical lens on agentic engineering — what schools figured out centuries ago that agent systems still need.
What I keep writing about
- 1Classrooms — why agents need practice environments separated from production.
- 2Textbooks — skills and docs as curriculum: one word, one meaning, no hidden prerequisites.
- 3Sleep — context compaction and memory consolidation as rest, not loss.
- 4Teachers — evals and feedback loops that make a quality bar enforceable.
School Core on Substack
Longer essays on building environments where agents — and the people who work with them — can actually learn. New posts as the school gets built.
schoolcore.substack.com ↗
✦ office hours
Chat with my agent
Curious what I'm up to, or sourcing me for a role? Ask away — human or agent, same desk.
✦ let's talk
Hiring for agent infrastructure?
Hiring for agent infrastructure?
You just found your person.
New York based, operating live systems in production, 10 merged upstream PRs across the MCP Rust SDK, OmniRoute, and Orca — and a teacher's instinct for making complex systems learnable.
Email →
brandonbennett@pursuit.org
GitHub →
github.com/branben · code, PRs, projects
LinkedIn →
linkedin.com/in/brandonebennett1
— Brandon
© 2026 Brandon Bennett · AI engineer, agent infrastructure · New York
Robots welcome — read llms.txt first 🤖