liveBook · Manning

Collaboration

As explained in Designing AI Agents

An agent capability for coordinating information or activity with other agents or humans.

Collaboration coordinates work across agents or humans pursuing a shared goal. Separate contexts can contribute bounded work, distinct expertise, or independent evidence, but their contributions need explicit handoff and integration contracts. Even agents using the same model need distinct responsibilities, evidence, tools, or criteria. section 2.2.1 section 8.1

Shared storage makes artifacts available; it does not automatically transfer decision state. Memory governs persistence and retrieval, while collaboration assigns tasks and authority, communicates results, and determines who accepts the integrated work. A delegated worker may therefore need versioned findings and source references rather than its parent’s entire conversation or tool history. section 4.7 section 8.1

Before adding another agent, pass three gates:

  • Establish a reason to split: independent search areas, different tools, separate authority, or a claim worth challenging independently. A large prompt alone is not enough. section 8.1
  • Define boundaries: name each participant’s artifact or question, read and write permissions, required return, and acceptance owner. section 8.1
  • Measure the advantage against a credible single-agent baseline. Compare quality, coverage, latency, tokens, retries, failures, and integration work. section 8.1
Table 8.2 Four collaboration patterns
Pattern
Coordinate
How responsibility is divided
Pressure point
Handoff Chain
Collaboration × Chain
Successive specialist phases
A weak or incomplete handoff travels downstream
Fan-Out/Gather
Collaboration × Parallel
Independent workers return to one gather point
Partition and merge
Adversarial Review
Collaboration × Loop
Participants test opposing sides of one claim
Independent evidence and a bounded verdict
Hierarchical Delegation
Collaboration × Hierarchy
Manager owns the goal; workers own subtasks
Manager framing and integration

Compare four ways to divide responsibility and the pressure point each introduces: handoff, partition and merge, independent evidence, or manager integration. section 8.1

Independent rollout check

For a controlled-rollout change, collaboration can give reflection an independent point of view. Governance remains a separate responsibility: it defines who may increase exposure, what evidence permits the next step, and how to stop the rollout. section 2.1.1

More agents do not guarantee better results. A team can become slower, more expensive, and less reliable. In one reported research evaluation, a multi-agent system used about 15 times the tokens of an ordinary chat while achieving a 90.2% lift over single-agent Opus 4; that trade-off depends on the task and evaluation. section 8.1

How Designing AI Agents covers it

How deep it goes
Sets out three adoption gates and four responsibility-splitting patterns without detailing their implementations.
Examples
Illustrates independent checking in a controlled rollout and reports a multi-agent research evaluation with higher token use.
What you can do after
Evaluate whether to split a task and compare a multi-agent approach with a single-agent baseline.
How it connects
Distinguishes collaboration from memory, reflection, and governance while linking their responsibilities.

Also explained in

Compare how 9 books explain collaboration →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.