I lead AI adoption at a listed company — in practice that means I'm the whole team: strategy, build, rollout, and the awkward conversations with the security reviewers. A fair amount of it I'm still figuring out as I go.
What I keep coming back to is the agent harness: the layer around the model rather than the model itself. Scheduling, memory, skills, evaluation, guardrails.
My working hypothesis, still being tested, is that model capability stopped being the bottleneck a while ago — and that what holds adoption back inside a company isn't the model not being smart enough, but the layer around it not being reliable enough yet.
- Getting agents into environments that hold secrets. Sandbox boundaries, permission matrices, approval interrupts, and the paths that quietly escalate privilege.
- Making token cost predictable. Tiered routing, context compression, and reasoning budget allocated by task complexity.
- Letting legacy systems absorb AI. Twenty years of process and data don't get a rewrite. The agent has to meet them where they already are.
- Making outcomes measurable. Without evals there's nothing to iterate against, and as far as I can tell that's where most enterprise AI efforts quietly stall.
I contribute mostly around guard boundaries and data integrity — less "add a feature", more "where can this be bypassed, where does it hang, where does it drop data without saying so". It's slow work that doesn't demo well, but it's the part I find most interesting, and I learn a lot from the reviews.
| Area | |
|---|---|
| Command execution guards | privilege and wrapper prefixes, path handling, option parsing |
| Checkpoint & rollback | making "restored" actually mean restored |
| Sandbox & network policy | proxy interference, honest network posture reporting |
| Gateway access control | isolation and addressing across profiles |
| Runtime resilience | idle timeouts, session state reclamation, health checks |
| Config & CLI | lossless parsing, consistent key validation |
| Web UI & i18n | rendering, form state, integration tests, locale catalogs |
OpenSquilla — token-efficient microkernel agent. A local router sends each turn to the cheapest model that is good enough, which maps directly onto the constraint I keep running into at work. → my pull requests
Hermes Agent — self-improving agent with a built-in learning loop; it generates skills out of its own usage and keeps iterating on itself. → my pull requests
Also reach for React, Tailwind, Swift, SQLite, pandas, NumPy, Vercel, Stripe, Power BI and Streamlit depending on what the problem needs.
Open an issue or a discussion. Always happy to compare notes on harness design, evals, or getting an agent past a security review — I learn as much from other people's answers as from my own.