Multi-agent systems

Why use more than one agent?

Start with a warning: most tasks need one good agent. Adding agents adds cost, delay and more places for things to go wrong. But there is a point where a single agent stops coping. Its instructions grow to cover five jobs, it has forty tools and picks the wrong one, and its context fills with material that is irrelevant to the step it is on.

At that point, splitting the work helps. A focused agent with one job, a short set of instructions and six tools is more reliable than a generalist with everything. The question becomes how those focused agents should work together. There are four basic shapes.

The four shapes

• Single agent. One agent, one loop. Simple to build, simple to reason about, simple to audit. The right default.
• Vertical (hierarchical). A lead agent breaks the task down, delegates pieces to specialist agents, and combines their results. Clear control and clear accountability. The weakness: the leader is a bottleneck and a single point of failure. If it misunderstands the task or splits it badly, every specialist does the wrong work well.
• Horizontal. Peer agents work as equals, passing work and critiques between them, for example a drafter and a reviewer. More flexible and good for debate or review, but coordination is slow, runs are harder to predict, and it is harder to say who decided what.
• Hybrid. A lead agent at the top, with small peer groups beneath it. Most large systems end up here: hierarchy for control, peers where review or debate improves quality.

Delegation is a contract

In a hierarchical system, the quality of the result depends on the quality of the hand-off. When a lead agent delegates, the request to the specialist should read like a good brief to a colleague:

• Purpose: what this piece is for and how it fits the whole.
• Inputs: the facts the specialist needs, so it does not have to guess or re-fetch.
• Expected output: format, length, level of detail.
• Limits: what it must not do, and when to stop.

Two governance rules matter just as much. First, a specialist never gets more rights than the person who started the work. Delegating to another agent must not become a way around approvals or access limits. Second, delegation needs a depth limit. Agents that can delegate to agents that can delegate again can loop, multiply cost, and bury accountability three levels down.

Common failure modes

• The telephone game. Each hand-off loses detail. By the third agent, "draft for internal review" has become "publish."
• Duplicated work. Two specialists both fetch the same data, or both make the same change.
• Cost multiplication. Every agent reads its own context. Five agents can cost far more than one.
• Lost accountability. When something goes wrong, you need to know which agent did what, on whose behalf. Each delegated run must leave its own record.

Checking a multi-agent result

A combined answer can look polished while hiding a weak link. Before you rely on one, check it the way you would check a report assembled by several colleagues. First, trace each important claim back to the specialist that produced it: a figure in the summary should match the figure in the finance agent's table. Second, read the briefs the lead sent, not just the answers that came back; most errors start in a vague brief. Third, look for gaps between pieces: nobody owned the part of the task that fell between two specialists. Fourth, compare the cost with what a single agent would have spent. If the multi-agent version is not clearly better, the simpler design wins.

When not to go multi-agent

Before splitting, try the cheaper fixes: tighter instructions, fewer tools, a better plan (lesson 3). If the task has fixed steps that never change, a workflow with one agent per step (lesson 7) is more predictable than agents delegating freely. Reach for dynamic delegation when the work genuinely needs different expertise and the split cannot be known in advance.

Multi-agent work in Renly

Renly's default agent, named Renly, can consult specialist agents from your organization for part of a task, then combine what they return. Each consultation runs as its own conversation with that specialist, on behalf of the same person, and delegation is limited to three levels deep. For fixed multi-step work, automations put a different agent on each step, which is the hierarchical shape with the plan written down in advance.