Multiplayer Engineering

I built HighRoad’s AI platform alone and had the first production version live in 60 days in Fall 2025. I barely wrote code myself. Instead, I directed and corrected AI agents. Since then, what’s possible for agentic engineering has changed considerably. I evolved a number of other add-ons that made shipping software faster and faster: sales conversations captured in Granola became structured PRDs in Linear, planning agents scoped work across repos, implementation agents built it using test-driven development, all against a shared set of repo standards and integrations.

Every engineering team now uses AI, but many CEOs don’t see it in their software output. Adoption stopped at the individual. AI arrived as a personal tool, each engineer found their own way to use it, and nothing about how the team builds, tests, deploys, or secures its work changed to match.

• AI is adopted one engineer at a time. Each person crafts their own prompts, skills, and workflows, and none of it is shared.
• The development process never adapted. Testing, review, and deployment run as they did before AI. Deployment is a bottleneck as code production increases.
• An agent cannot be trusted to act end to end with open permissioning. Without a systems-level view of what an agent may touch, it stays a supervised assistant rather than something that can work a problem on its own.
• Code is generated locally on laptops, so product leaders and non-engineers can neither see what is being built nor try a contribution of their own, and agents only run while someone’s machine is open.
• Experimentation has stalled. Background agents, multi-model setups, and role-based agents have not been tried in months.

Making engineering multiplayer

The fix is to treat the way the team builds as a system to design, not a tool each person picks up on their own. That means standardizing the ground everyone builds on, changing the process to match AI’s output, and opening the work so it runs beyond any one laptop.

• Align on a set of repo standards and a map of the codebase, then encode them as shared repo skills every engineer’s AI draws from, so one person’s improvement propagates to the whole team.
• Rebuild testing and deployment around AI’s volume, with stronger automated tests and evals, agent-assisted review, and CI that can absorb the throughput instead of gating it.
• Design permissioning from a systems perspective, with scoped credentials, sandboxes, and guardrails, so agents can act autonomously inside safe boundaries instead of waiting on a human at every step.
• Enable cloud agents to work alongside local, so the work is visible, product leaders can contribute directly, and agents keep running when no one is at their desk.
• Run a standing cadence of experimentation across background agents, multi-model routing, and role-based agents, so the build system keeps improving as the frontier does rather than freezing at first adoption.