Agent Coordination Structure

In my day-job, I'm a leader of Engineering Practice Improvement. I help teams build better software - faster. We can talk more about my feelings on all of that soon¹.

I've also been vibe-coding since March with the rest of you.

I'm impatiently waiting for tools and workflows to catch up to the new pace of development.

So What Are We Gonna Do About It?

We are going to organize a system of work that allows us to:

• give feedback at the right moment.
• trust our agents at the right moment.
• build faster and better.

In copilot-personas, I leverage three mechanisms to coordinate - reducing cognitive load for myself and our agents.

A familiar challenge in this framework is the balance of constraint and autonomy. Each token adds direction - but may incur noise and cost. Effective Context Engineering leaves processing space to solve the problem.

We want to ensure Just-In-Time (JIT) knowledge acquisition with clear process and artifacts.

Critical Directives

Tightly scoped universals reduce ambiguity. By defining clear boundaries, we stop the agent from guessing and let it focus on execution.

• Context-Aware Guidance: Instructions that auto-activate based on the file type you're working on.
• Narrow Tech Standards: Explicitly define versions and libraries (e.g., "React 18 with TypeScript") to prevent hallucinated dependencies.
• Golden Paths: Pre-defined architectural patterns that guide the agent down the "happy path" of our codebase.

Artifact Handoffs

Required artifacts encapsulate context. Recent tooling leans into Plan and Execute modes facilitated by document-based handoffs.

• Encapsulated Context: Specific artifacts (like a PRD_[feat].md) that contain everything needed to define our next step.
• Structured Transitions: Moving from Persona A to Persona B using VS Code Handoffs to guide the workflow.
• Standardized Naming: Using consistent paths like ./artifacts/ so every agent knows exactly where to look.

Agentic Purpose

Specialized personas ensure reliable output. One agent can't hold all of our context, so we run a team of specialists.

• Specialized Personas: Agents with archetypal roles (Architect, Builder, QA) rather than a generic "Assistant".
• Modular Skills: Reusable capabilities (like "PDF processing") that load on demand, keeping the context light.
• Context Economy: Treating context as a scarce resource. We only load what's needed for the specific task to prevent confusion.

But why?

Soon, industry tooling standards and model capabilities will catch up. I already told you that I'm impatient.

Fortunately, this has been an exercise in context and philosophy. What do I care about, and how do I deliver that message efficiently?