Summary
Craig Hewitt shares 25 hard-won lessons from a month spent completely rebuilding the Hermes agent that runs his seven-figure SaaS business. Unlike theoretical AI tutorials, this video covers real operational decisions: how to define an agent’s job before building, how to give it an identity (his is named Agent Smith, with sub-agents for sales, marketing, and product), how to write a soul.md file that makes the model warmer to interact with, and how to structure context access across read-only and read-write knowledge bases.
On the technical side, Hewitt uses GPT 5.5 as the underlying model and MCP as the primary integration layer — including the ChartMogul MCP for real-time SaaS metrics and Cursor for autonomous code generation. He walks through a Git-branching strategy for agents: new personalities and sub-agents are tested on a staging branch before merging into main, protecting mission-critical production workflows from experimental changes. Permission escalation is treated as an earned trust relationship, not a day-one grant.
Perhaps the most practically useful section covers team adoption: Hewitt argues that resistance comes primarily from context starvation (the agent gives bad answers because it lacks the information it needs) and from founders who fail to reframe the agent as amplification rather than replacement. He separates agent communication into dedicated Slack channels per sub-agent to reduce context switching, an approach he believes also benefits the model’s own coherence across long sessions.
📺 Source: Craig Hewitt · Published September 10, 2026
🏷️ Format: Workflow Case Study
