Production Agents
I build the tools, guardrails, and integrations that move AI models beyond chat and into production — where mistakes are expensive.
On our own servers, with Hermes, the open-source agent runtime by Nous Research.
One company,
nine agents
One control plane and eight department profiles, each one serving a real team with a human owner. People talk to them, and nothing merges without a person saying so.
leadership & advisorLeadership6 with access
control planeOrchestrator12 with access
profileEngineering11 with access
profileSales17 with access
profileMarketing13 with access
profileKey Account3 with access
profileProjects5 with access
profileCS9 with access
profileSupport21 with accessOf the company uses it regularly, almost all outside tech leadership.
PRs merged in five months, across 35 repos, from product code to the company site.
AWS cost cut through automated resource & disk pruning.
Token spend ceiling per profile, per month.
Inside the tools
Each agent works in the software its team already uses. Nobody switches tools to use it.
Triages incoming tickets every 10 minutes during business hours.
A morning pipeline briefing and the timeline of every account.
CS owns the product, so its agent reads how each customer uses Pantera.
Ships site changes, measures leads in GA4, routes each lead to the right salesperson.
Turns issues into PRs, triages CI, watches cloud cost.
Alerts, fleet users, and the briefings that cross departments.
48 scheduled routines across the fleet
Guardrails
Production repositories need more than a good model. These make delegation real.
- 01 // Human go-ahead
Nothing merges on its own
No direct pushes, and at most 2 open PRs per repo. In production code, an engineer reviews and merges. In internal tools, like the company site, the team that owns it gives the go-ahead in chat and the agent merges.
- 02 // Cost & safety
Scripts before LLMs
Anything that can be solved with Bash, Python, or Git never calls an LLM. This prevents hallucinations, runs quiet security checks, and keeps token spend strictly under $140/mo per profile.
- 03 // Isolation
Each agent in its own lane
Each profile runs in an isolated sandbox with its own tools and procedures. The sales agent never touches source code, and context never bleeds across domains.
- 04 // Real scope
Work from the backlog
Issues the team actually opened: async race conditions, component migrations, end-to-end test suites, lead routing into the CRM.
Self-hosted
End to end — the agent runtime, SQLite state, and every integration run on our own infrastructure, decoupled from inference providers. Swapping the model is a config change; the guardrails and review process stay the same.
Write to me
If you're building something ambitious with agentic systems, write to me.