OpsAgent is an internal-tool / workflow builder for ops, finance, and customer-success teams. Think Zapier with first-class LLM nodes, evals, observability, and a rollback plan baked in.

A 220-person SaaS company had 6 ops engineers maintaining ~80 brittle Zapier + Retool flows. Half of them silently failed once a month. They wanted a tool where non-engineers could build durable AI workflows without spawning yet another shadow-IT problem.

A visual workflow canvas with triggers, actions, AI steps, conditionals, and parallel branches. All stateful, all retriable.

60+ pre-built integrations (HubSpot, Salesforce, Stripe, Slack, Notion, Postgres, GitHub, Zendesk).

AI step types: extract, classify, summarize, decide, route, write-with-tools. Each with versioned prompts, eval traces, and human-in-the-loop checkpoints.

A run-explorer with full step-by-step traces, cost per run, and one-click rollback to a prior version.

Workspace + role-based publishing (sandbox, staging, production).

Frontend: Next.js + React Flow + Tailwind, Monaco editor for prompt + JS expressions

Backend: Python (FastAPI) + Postgres + Temporal.io for durable workflow execution

AI: OpenAI + Anthropic + open models behind an internal LLM gateway with caching, rate limiting, and per-workflow cost tracking

Eval framework: golden test cases per workflow + LLM-as-judge for nuanced asserts

Infra: AWS (EKS), Datadog, ClickHouse for telemetry

Replaced ~70% of internal Zapier/Retool flows in 5 months.

Failed-run rate dropped from 8% to 0.4% post-evals.

Ops engineers freed up to focus on truly bespoke automations; non-eng folks ship workflows daily.