Behind the Build: What is an Agent Harness? — with Kinan Hamwi (Cerevisor)

Kinan Hamwi breaks down what an "agent harness" actually is (the system that turns a raw model into something that can act on your computer), then demos Cerevisor, his own local-first harness that runs Claude, Codex, Gemini and other models side by side, tracks skills, memory and analytics per agent, and keeps everything owned on the user's machine. 00:00 Intro: what is an agent harness, and why now 01:20 Kinan introduces himself and Cerevisor 03:34 Model vs. harness: the chef-and-kitchen analogy 05:11 What a harness tracks: context, memory, tools, task state, permissions, recovery 08:25 The harness as the "digital body" of the model 08:53 Maturity ladder: chatbots to RAG to tool use to agents to harnesses to enterprise 12:44 Q&A: Anthropic's Claude-in-Slack ("cloud tag") as an enterprise harness attempt 15:12 Cerevisor demo begins: the workflow canvas 16:00 Experiment: same prompt across Gemini Antigravity, Claude, GPT-5.4, GPT-5.5 17:55 Assigning and removing skills per agent, per model 23:24 Built-in secrets vault, replacing scattered .env files 24:45 Live build: a content-pipeline workflow (Kat's "AI for real estate" use case) 29:22 Pricing and availability: free now, paid tier coming with community pricing 31:02 Reviewing the generated workflow: not every skill works with every model 37:32 Loop versioning: rerun as a loop, inspect every iteration, roll back versions 38:59 Harness memory: the "self-portrait", fully local and exportable as JSON 40:44 Analytics: run history, best model per task 41:10 Q&A: how a Claude subscription's own memory relates to Cerevisor's memory 42:31 Reviewing the output: the generated research brief and PowerPoint deck 45:22 Roadmap: testing GLM 5.2, aiming for full independence from cloud providers 45:56 Wrap: build timeline, and a teaser for a future session on managing a live repo while agents run Recorded as part of the Happy Operators community call series (Behind the Build / AI & The Rest).