hawkes.aiKey ideas
- The one idea — One agent hits limits — finite/rotting context, one perspective, one skillset — so complex work benefits from multiple specialized agents that communicate: dividing labor, each with fresh context and focus, coordinating toward a goal.
- How it is shown — A single overloaded agent (context rotting) versus a team — researcher, coder, reviewer — each with fresh context, passing results to each other via an orchestrator.
- The trap to avoid — Assuming one big agent is always best — for complex, long, multi-part work, one agent's context and focus become bottlenecks (but more agents isn't always better — G79).
- What it sets up — Before orchestrating many agents, understand the wrapper that makes even one work: the harness.
Why would one AI need to talk to another? Because a single agent has one context that fills and rots, one skillset, one focus. Big messy work needs a team.
The one idea
One agent hits limits — finite/rotting context, one perspective, one skillset — so complex work benefits from multiple specialized agents that communicate: dividing labor, each with fresh context and focus, coordinating toward a goal.
One agent is powerful — but it hits walls. Its context fills and rots. It carries one perspective, one set of tools, one focus. For big, messy, multi-part work, a single agent becomes the bottleneck. The fix looks like a human team: multiple agents, talking. Start with context. A single agent's window is finite, and over a long task it rots — the goal buried, the signal drowned. Split the work across several agents, and each gets its own fresh, focused context, none drowning in the others' history. Then specialization.
How it works — the demo
A single overloaded agent (context rotting) versus a team — researcher, coder, reviewer — each with fresh context, passing results to each other via an orchestrator.
Instead of one generalist juggling everything, you build a researcher, a coder, a reviewer — each with its own tools and prompt, each doing one thing well. It's a lone jack-of-all-trades versus a team of experts. But specialists only help if they coordinate. So agents pass results and messages to each other: an orchestrator delegates a subtask, a worker does it and reports back, another checks it. Work flows between them, like a team handing off pieces of a project. And that coordination needs a way to communicate — the same standardization drive behind MCP, now extending agent to agent. MCP standardized model-to-tool; the same push is standardizing how agents connect and hand off work. The ecosystem logic, one level up. The trap: assuming one big agent is always best.
The trap to avoid
Assuming one big agent is always best — for complex, long, multi-part work, one agent's context and focus become bottlenecks (but more agents isn't always better — G79).
Why it matters — and what’s next
Before orchestrating many agents, understand the wrapper that makes even one work: the harness.
For a quick task, it is — simpler and cheaper. But for complex, long, multi-part work, one agent's context and focus become the ceiling, and a coordinated team can break through. Sometimes. Not always. So one agent hits limits — context, perspective, focus — and teams of communicating specialists go further. But before you orchestrate many agents, understand the thing that wraps even one model into a working agent. What is that wrapper? Next group: the harness.
This is one short episode in AI: Zero → Frontier, a step-by-step climb through how AI actually works. Each episode builds only on the ones before it.
Full transcript 2:30 of narration
One agent is powerful — but it hits walls. Its context fills and rots. It carries one perspective, one set of tools, one focus. For big, messy, multi-part work, a single agent becomes the bottleneck. The fix looks like a human team: multiple agents, talking.
Start with context. A single agent's window is finite, and over a long task it rots — the goal buried, the signal drowned. Split the work across several agents, and each gets its own fresh, focused context, none drowning in the others' history.
Then specialization. Instead of one generalist juggling everything, you build a researcher, a coder, a reviewer — each with its own tools and prompt, each doing one thing well. It's a lone jack-of-all-trades versus a team of experts.
But specialists only help if they coordinate. So agents pass results and messages to each other: an orchestrator delegates a subtask, a worker does it and reports back, another checks it. Work flows between them, like a team handing off pieces of a project.
And that coordination needs a way to communicate — the same standardization drive behind MCP, now extending agent to agent. MCP standardized model-to-tool; the same push is standardizing how agents connect and hand off work. The ecosystem logic, one level up.
The trap: assuming one big agent is always best. For a quick task, it is — simpler and cheaper. But for complex, long, multi-part work, one agent's context and focus become the ceiling, and a coordinated team can break through. Sometimes. Not always.
So one agent hits limits — context, perspective, focus — and teams of communicating specialists go further. But before you orchestrate many agents, understand the thing that wraps even one model into a working agent. What is that wrapper? Next group: the harness.