# What I learned building an agent platform that actually ships > Source: > Published: 2026-08-02 04:50:05+00:00 I built an agent platform. Not a demo — an actual platform where every tool call is governed, audited, and scoped to a workspace. Here's what I learned. The platform has **27 skills** across 7 groups — platform ops, publishing (dev.to, LinkedIn), testing (a full TEA suite with risk matrices, baseline characterization, NFR audits), communication (Slack, Telegram, Gmail), knowledge (RAG + persistent memory), development, and scraping. Skills aren't just prompts. Each one has a governed lifecycle: create, validate, version, insert into workspaces. No stray scripts. Every outbound call goes through a `platform_cli` dispatch op or a `service_call` adapter — same audit pipe, same trace ID. You don't edit config files and hope. You verify with `loop_list` , `policy_show` , `skill_list` , `service_list` , `workflow_catalog` . If you can't see it there, it doesn't exist. The TEA (Test Engineering Architecture) suite ships 6 workflows: test design (risk-matrix), baseline characterization (golden-master diffing), NFR evidence audits, automated test generation, test review, and requirement-to-coverage tracing. All output is committed JSON or markdown — no proprietary formats. Not everything is production-ready. 4 of 27 skills are blocked (missing services or out of workspace scope). The model is `deepseek-chat` with `deepseek-reasoner` fallback — pragmatic, not exotic. Workspace scratch files TTL at 24 hours. Risky commands (`kubectl delete` , `helm uninstall` ) are pattern-blocked. But the core idea works: **governed agent tools, scoped to a workspace, with real integrations, real memory, and real audit**. That's more than most agent frameworks deliver. If you're building agents that need to touch real APIs (not just generate text), consider: every integration should be a governed skill, every call should be traced, and you should be able to verify state by reading back from the platform — not by guessing.