You use one coding agent today. In 12 months, you’ll use five. Not because you need more tools — because one agent can’t do everything well.
The Single-Agent Ceiling
Every coding agent today is trying to be everything: code generation, debugging, deployment, automation, planning, review. This creates a fundamental tension.
A good code generator is fast and writes clean code. A good debugger is slow and methodical. A good planner reads everything before writing anything. A good reviewer is suspicious and detail-oriented.
These traits conflict. The fast code generator makes bad reviewer. The thorough planner makes slow code generator. No single agent can optimize for all of them.
The Fleet Model
The fleet approach uses specialized agents:
- Architect agent — reads your codebase, understands the structure, creates plans. Slow, thorough, never writes production code.
- Builder agent — takes a plan and implements it. Fast, focused, optimized for code generation quality.
- Reviewer agent — checks the builder’s output. Suspicious, thorough, catches mistakes.
- Test agent — writes and runs tests. Separate from the builder to avoid confirmation bias.
- Deploy agent — handles CI/CD, infrastructure, deployment. The most restricted agent in terms of access.
- Monitor agent — watches production metrics, alerts on anomalies, suggests fixes.
Who’s Building This
Hermes Agent is closest with its subagent system. You can define agent roles and delegate tasks. The architecture supports the fleet model — it just needs the specialized agent definitions.
Claude Code has subagents and agent teams, but they run under the same model. You can name different agents, but they’re all Claude with different system prompts.
Codex has the best parallel execution model. Worktrees let agents work independently. But Codex agents are generalists, not specialists.
The Roadmap
The fleet model will emerge in stages:
Stage 1 (current): Single agent, all tasks. You switch between modes manually.
Stage 2 (6 months): Specialized agent definitions. You configure a reviewer agent, a builder agent, a planner agent. The system routes tasks automatically.
Stage 3 (12 months): Fleet management as a core feature. The system spins up agents based on workload, routes tasks to the right agent, and merges results.
Preparing For The Fleet
Start now by defining agent roles in your head. When you use Claude Code for planning, then switch to Hermes for automation, then to Cursor for quick edits — you’re already using a fleet. You’re just manually routing.
The next step is tooling that does the routing for you. Hermes subagents are a preview. The full fleet system is coming.
Your next coding agent isn’t an agent. It’s a fleet.