🧑🤝🧑 Hermes as a team
Instead of doing everything in a single thread of reasoning, Hermes can become a orchestrator: it divides the work among assistants, each focused on a subtask. It’s the same idea as a manager assigning tasks to the team—instead of doing everything alone.
🎼 Orchestrator + executors
The orchestrator doesn’t do the heavy lifting—it plans, delegates, and brings it all togetherSub-agents handle the execution. Large tasks become manageable when they’re broken into smaller ones that run at the same time.
🧠 Fresh context for each agent
Each sub-agent is created with its own its own clean context window, dedicated only to your subtask. This is important: remember Track 2 — the fuller the context, the worse the performance. Giving each agent lean context means better, cheaper answers.
📊 Why isolated context wins
- Focus — the agent isn’t distracted by the history of other tasks.
- Cost — less context = fewer tokens per call.
- Quality — with no noise, the answer is more precise.
⚡ Parallel, not sequential
Here's the turning point. In a sequential workflow, the agent does task A, then B, then C — the sum of the times. In the in parallel, 4 to 6 agents work at the same time and deliver together at the end. Instead of 6 hours in a queue, 1 hour working simultaneously.
✗ Sequential (slow)
- ✗Task A (1h) → B (1h) → C (1h)…
- ✗Total time = the sum of all tasks.
- ✗6 tasks ≈ 6 hours.
✓ Parallel (fast)
- ✓A, B, C… all at the same time.
- ✓Total time ≈ the longest task.
- ✓6 tasks ≈ 1 hour.
💡 Practical tip
Parallelism only works when the tasks are independentIf step C depends on the result of B, they can’t run at the same time — sequential execution is unavoidable.
🌎 Example: job search
The classic example from the material: "research the best AI companies to work for"Hermes comes up 2 sub-agents — one covers the US, the other the international market. Each researches in parallel, and in the end both hand off to the orchestrator to compile the report.
How a request becomes a team (illustrative)
🏭 12 parallel instances: the co-founder case
Parallelism doesn't stop at 2 or 3. The Hermes co-founder runs 12 parallel instances every day to build Hermes itself — monitoring issues, dogfooding, and managing the kanban. It’s a fleet of agents working like an entire engineering team.
Monitor issues
Agents track bugs and reported requests, triaging incoming items.
Dogfooding
Use the product itself to find problems before users do.
Manage the kanban
Move tasks, update status, and keep the workflow running.
📊 The number that matters
12 instances running in parallel, every day. Parallelization isn’t a demo trick — it’s how power users actually work.
🎭 Sub-agent roles
As on a real team, each agent can have a role: research, writing, design, scheduler. Clear roles prevent two agents from doing the same thing—and each role can use the ideal model (expensive reasoning only where it’s worth it; volume on the cheap one).
Collects and cross-references data.
Writes the deliverable.
Formats and presents.
Coordinates deadlines.
⚖️ When to parallelize (and when not to)
Building a team adds overhead: each sub-agent incurs the ~73% fixed cost of the request (module 3.5). For a small task, the orchestrator + 5 agents may cost more than it’s worth. Parallelize when the time saved justifies the extra cost.
✓ Worth parallelizing
- ✓Independent, time-consuming tasks (broad research).
- ✓Clear filters (region, topic, source).
- ✓When 6h in a queue becomes ~1h in parallel.
✗ Not worth it
- ✗A small task that one agent can solve quickly.
- ✗Dependent steps (C needs the result of B).
- ✗When the overhead of 6 agents outweighs the benefit.
💡 Practical tip
Before bringing up a team, ask: "can these tasks run at the same time, and is the time saved worth the overhead?" If so, parallelize; if not, stick with a single agent.
📌 Module Summary
Next Module:
3.4 - 💓 Heartbeat / Cron: the heartbeat that keeps the agent alive and working 24/7.