I run AI agent swarms in production. Clear roles, written procedures, QA gates in code, and hard limits on what runs without me. They do in hours what used to take a team months, inside the ad account, the CRM, and the phone line.
Me on top: direction, review, and everything that spends money or can't be undone. Below me, a swarm: many agents in parallel, one job each. The task picks the models: frontier for reasoning and code, small and fast for proven actions, local when data shouldn't leave the machine. The harness is everything you'd hand a new hire, written down. A separate policy gate checks every action, masks secrets, and keeps untrusted content from taking control. A model scores the work; a threshold in code makes the call.
- Voice agents in production: self-hosted voice platform operated end to end from the terminal and by agents over MCP (dograh-ctl: 33 commands, 25 MCP tools, on PyPI). Real calls, transcripts, latency budgets.
- Multi-tenant voice and automation platform: per-tenant Postgres row-level security, prepaid wallets with billing that stays consistent under concurrent webhooks, and a two-stage call-classification pipeline (deterministic rules plus an LLM tie-breaker) at 14/14 on its eval set.
- Autonomous pipelines: ingest, score, act on a schedule (market monitoring, lead generation, ad creative), each with run-cost budgets as first-class constraints.
- Agents own outcomes, not steps. If it can't own an outcome, it's not ready.
- LLM-as-judge everywhere, but the threshold lives in code: testable, versioned, immune to model mood.
- Every mistake becomes a rule the next session follows.
dev@erickpaniagua.com · erickpaniagua.com · Detroit, remote (US)