Harnesses for AI agents: Claude Code, Codex, Cursor, and turning agent mistakes into rules of its environment.
How an agent’s commands (control plane) differ from the data it reads (data plane), how prompt injection works and how to defend: least rights, approval of writes, separating read from write, trusted sources.
How to investigate an incident with an AI agent: connect Sentry and Grafana through MCP, have it match errors, metrics and release time, limit permissions and verify the hypothesis.
The three MCP server scopes in Claude Code: local, project and user. Where the setup lives, who sees it, what wins when names collide, how to share servers with a team and keep secrets out of commits.
What to demand from an agent in requirements and meeting notes: a source link, the model’s assumptions and a “fact or not a fact” status. Where major and minor edits come from and how they relate to use cases.
We optimise one step of the chain: QA after support. An agent reproduces a Jira bug on a test stand via MCP, narrows the path, describes and documents it, a skill makes the steps a repeatable process, QA checks by hand. Support and development do not change.
MCP plugs an agent into GitLab, Jira, Confluence and meeting recordings. What problem it solves, how to set it up in minutes and delegate that to the agent itself, how it saves time and tokens and why the agent invents less.