Skip to main content

Building Effective Agents (and Teams of Agents)

Building a great agent starts with a clear goal and a well-defined scope. Whether you're creating a single research assistant or a team of specialized agents that pass work between each other, the same principles apply: give agents focused instructions, the right tools, and clear boundaries.

Designing a Single Agent​

Start with a Clear Purpose​

Every agent should have one primary job. Instead of creating a single agent that tries to "handle everything," create focused agents:

❌ Too Broad✅ Focused
"Monitor competitors and send reports and also manage billing""Monitor competitor pricing changes and alert when prices drop"
"Research topics and reply to customer emails and track shipments""Research industry trends and compile weekly briefings"

A broad scope can make results and inference usage harder to evaluate. Start with a bounded responsibility and compare actual results before expanding it.

Write Great Instructions​

The agent's charter supplies its standing instructions. Result quality also depends on its model, context, tools, permissions, and connected services.

Do:

  • Use clear, direct language: "Send a daily summary of trending topics to the team Slack channel."
  • Include specific output formats: "Format results as a Markdown table with columns: Topic, Source, Relevance."
  • Set boundaries: "Only research topics related to artificial intelligence and machine learning."
  • Describe edge cases: "If no new results are found, skip today's report — don't send an empty message."

Don't:

  • Use vague language: "Keep me updated on stuff."
  • Give multiple unrelated responsibilities.
  • Forget to specify what not to do.

Set the Right Intelligence Level​

Choose from the intelligence settings available in your deployment. Test a representative task and compare correctness, required corrections, elapsed time, and inference usage per useful result. A stronger tier does not guarantee a better or faster outcome, and a cheaper cycle can require more retries.

Watch Usage​

Open Usage to compare measured requests and tokens with the results an agent produced. If your administrator has enabled usage limits, a limit can block additional work; neither a limit nor an in-flight check guarantees exact spend.

  • Start narrow — give new agents a bounded responsibility and a small first task
  • Monitor usage — check Usage to see actual consumption
  • Expand deliberately — widen scope only when the agent is productively using what it has

Building Teams of Agents​

When your workflow involves multiple steps or domains, consider splitting work across several specialized agents connected via peer linking.

When to Use Multiple Agents​

Use a Single Agent When...Use Multiple Agents When...
The task is straightforward and linearWork spans distinct domains (e.g., research → drafting → review)
A single instruction set covers everythingEach step needs different tools or permissions
You want the simplest possible setupYou need parallel work streams
Latency isn't criticalEach subtask is deep enough to justify its own agent

The Research → Draft → Review Pattern​

A common multi-agent setup follows a pipeline:

┌────────────┐     ┌────────────┐     ┌────────────┐
│  Research  │ ──▶ │   Writer   │ ──▶ │  Reviewer  │
│  Agent     │     │  Agent     │     │  Agent     │
│            │     │            │     │            │
│ Gathers    │     │ Synthesizes│     │ Fact-      │
│ sources &  │     │ findings   │     │ checks &   │
│ data       │     │ into draft │     │ edits      │
└────────────┘     └────────────┘     └────────────┘

Each agent has:

  • A focused instruction set
  • Different tools (research uses web search; writer uses email/docs)
  • A peer link to pass results to the next agent in the chain

Design Your Agent Team​

  1. Map the workflow — List the distinct steps from start to finish
  2. Assign one agent per step — Each agent owns one responsibility
  3. Define handoffs — What does each agent pass to the next? (e.g., "After researching, send a summary to the Writer agent")
  4. Set the trigger — Usually the first agent runs on a schedule; downstream agents run when they receive input from the previous agent
  5. Test the chain — Start with a manual trigger and verify each link before enabling the full schedule

Avoid Common Pitfalls​

  • Circular workflows — Don't create a loop where Agent A passes work to Agent B, which passes work back to Agent A. This consumes usage without progress.
  • Overlapping schedules — If two agents independently monitor the same source, they'll duplicate work. Assign clear boundaries.
  • Competing for contacts — When multiple agents share a contact (e.g., all sending to the same Slack channel), use templates to ensure consistent formatting and avoid conflicting messages.

Using Templates for Consistency​

Company templates let you reuse agent configurations within your company. Templates are especially useful for teams:

  • Standardize agent patterns — Create a "Research Agent" template that anyone in your org can instantiate
  • Share across workspaces — Save internal company templates so teams maintain consistent agent designs
  • Iterate on proven setups — When you find a great agent configuration, save it as a template

Next Steps​

  • Explore Peer Linking to connect your agents into teams
  • Learn about Planning and Deliverables to understand how agents show their work
  • Check Usage for broader guidance on platform budgets, caps, and operational limits
Last updated on