My Agentic Engineering Workflow (after 6,775 sessions)

David Ondrej demonstrates his multi-agent engineering setup, from interfaces and persistent remote sessions to reusable skills and production feedback. The durable theme is keeping decisions, delegation, and review deliberate rather than maximizing agent activity. This is a Hostinger-sponsored practitioner walkthrough, not a comparative productivity study; model rankings, subscription advice, and scaling predictions reflect his preferences at recording time.

Main takeaways

• Make agent state and priority visible. 02:57–04:50(opens in a new tab): Ondrej values interfaces that distinguish running, idle, and blocked agents. He also describes an unfinished priority-based interface intended to surface important completed work before lower-priority tasks, rather than interrupting the operator in completion order.
• Use remote sessions for work that must outlive the laptop connection. 10:04(opens in a new tab) onward: the VPS walkthrough combines SSH, Herder, development dependencies, and Pi; a running Pi session is shown at 32:01(opens in a new tab). Editorial caution: the setup uses root SSH access and authorizes the agent to find a laptop API key for reuse on the server 16:26(opens in a new tab), 22:20–23:17(opens in a new tab). Do not copy that access pattern as a default: use least-privileged execution and explicitly provisioned, narrowly scoped credentials. The demonstration does not establish isolation or tested recovery from disconnects.
• Package repetition, but keep consequential choices human. 27:05–32:01(opens in a new tab): short repeated instructions become text snippets; multi-step procedures become skills. His review skill combines two model reviews, while “ask then build” surfaces architectural and product choices before implementation rather than letting the agent silently decide.
• Separate parallel implementation from serialized shipping. 36:03–40:49(opens in a new tab): he describes an OS-level lock around merge, verification, push, CI, deployment, and health checks. Worktrees isolate concurrent edits; he avoids their overhead on small projects. His agent-count thresholds are personal heuristics, and a single-host lock is not evidence of coordination across multiple machines.
• Bound review and control delegation. 41:38–47:54(opens in a new tab): Ondrej recommends checking meaningful changes with a different model, fixing material findings, and avoiding recursive review loops that invent more work. He wants explicit control over when sub-agents run and which models they use. His direct-to-production examples are personal practice, not a general release policy.
• Persist the reasoning code cannot explain. 46:52–49:00(opens in a new tab): architecture decision records preserve why a design was chosen and its historical context so future agents and people can revisit it without reconstructing the conversation.
• Treat hooks and database roles as limited controls. 34:08–36:55(opens in a new tab): he describes pre-tool checks for selected dangerous command patterns. Editorial distinction: these are not isolation or a guarantee against destructive actions or prompt injection. At 49:58–51:52(opens in a new tab), he recommends read-only PostgreSQL roles for analyzing usage and errors before prioritizing features. Read-only permissions limit writes, not sensitive-data exposure; accessible data and what may be sent to models still require explicit scoping.

Full video: David Ondrej — My Agentic Engineering Workflow (after 6,775 sessions)(opens in a new tab).