A single agent can use many tools, load skills, keep state and pause for review. It can make multiple model calls. It can run independent tool requests in parallel. One agent does not mean one prompt or one uninterrupted conversation.
Start with the shape of the work
If the sequence is known, use a workflow. If the rules are exact, use code. If the next step needs judgment within one shared case, start with one agent. Consider a team when independent work or separate context has a measurable benefit.
When a team adds overhead
- A fast customer conversation. Extra starts and handoffs can delay a simple answer and make the voice inconsistent.
- Sequential work. If every step needs the previous result, parallel agents may have little to do.
- One shared history. Copying context costs tokens; summarizing it can lose useful details.
- Cheap, frequent tasks. Small coordination costs become material at volume.
- Shared records or tight API limits. Concurrent actions can conflict or create retries without increasing throughput.
- Limited maintenance capacity. More roles, prompts and handoffs mean more behavior to test and support.
When specialists earn their place
Independent research, bounded document batches and artifact checks can benefit from separate workers. A specialist may also help when a main agent’s context or tool selection is failing in measured tests. Parallel execution can lower elapsed time while increasing total compute. Measure both.
A separate agent is not automatically a security boundary. Access controls belong in the tools, data stores and execution environment. Two agents agreeing is not the same as two independent sources supporting an answer.
Keep the company’s operating context organized
Store company rules, department procedures, employee preferences and case history separately. Retrieve only what the task and the person’s permissions allow. A department can own a skill without needing its own agent.
Keep saved task state separate from a conversation summary. Recheck live facts before an action that changes a business record. Give memory an owner, correction process and expiry policy.
Compare designs on actual work
Use the same representative tasks, source access and acceptance rules. Include missing information, changing records and tool failures. Compare accepted completion, response time, human review, errors and maintenance—not just how impressive a demo looks.
Total cost per accepted completion = all operating costs across attempted runs ÷ accepted results. Include model calls, tools, infrastructure, review, rework and maintenance. Put implementation costs into a separate payback calculation. Recovered time is capacity; cash savings require a real business change.
What this means for Intelligence Factory
The $25,000 starting Workflow Build offer supports up to 30 agents where useful. That is capacity, not a target. We build and test the complete workflow. If one agent and a few reliable tools do the job better, that is the system we should deliver.
Try the architecture guide or explore the cost-per-result calculator.
Sources and further reading
This article combines provider guidance with Palnau’s implementation recommendations. The research is directional; it does not establish a guaranteed saving for a client.
- Anthropic: Building effective agents — workflows, agents and simpler starting points.
- Anthropic: When to fold subagents into the main loop — published October 22, 2026.
- Google Research: When and why agent systems work — benefits depend on the task.
- Anthropic: Context engineering — selecting and managing useful context.
- LangGraph: Persistence — saved state and reviewable execution.