Methodology · Agent Constructionemergingverified

Agent Count Escalation

also known as from-one-to-many, single-to-multi-agent escalation

Applies to: agentmulti-agent-system

Tags: multi-agentescalationcoordinationagent-count

Start with one agent. Add a second only when the first clearly fails the task and you can show it. When you do add agents, pick the coordination style on purpose: democratic, manager, hierarchical, or actor-critic. Do not reach for a swarm by reflex. Treat the number of agents as a dial you turn only when the evidence says to. Pair each step up with a coordination shape that fits the failure you saw.

Methodology process overview

flowchart TD base[Single-agent baseline metrics] --> s1[Build and measure single-agent baseline] budget[Coordination cost budget] --> s3 s1 --> dec{Single agent meets bar?} dec -- yes --> ship1[Ship single agent] dec -- no --> s2[Identify specific failure escalation should fix] s2 --> s3[Pick coordination style] s3 --> s4[Add smallest number of agents to resolve failure] s4 --> s5[Re-measure against same eval set] s5 --> dec2{Gains material?} dec2 -- yes --> s6[Document escalation as decision record] dec2 -- no --> rb[Roll back to baseline] s6 --> out1[Agent-count decision] s6 --> out2[Coordination shape]

Intent. Make 'how many agents' a decision driven by evidence, and force a deliberate choice of coordination style at each step up, instead of jumping to multi-agent by reflex.

When to apply. Use this when a team wants to design a multi-agent system before a single agent even exists, or when one agent has stopped improving and the next move is unclear. Don't apply it when the task is multi-actor by nature, such as negotiation, simulation, or debate as a feature. There the multi-agent shape is part of the requirement. Skip it too when you have no measurement of how the single agent fails, because then you have nothing to base a step up on.

Example scenario

A small ML platform team at a fintech was pushed by leadership to 'go multi-agent' for their compliance-document-review agent. They held out long enough to apply the methodology. Inputs: a single-agent baseline already in production, hitting 78% F1 on the compliance-eval set with median latency 12s per document, and a coordination cost budget of +$0.05 per document and +20s latency. The failures clustered in two places. First, the single agent missed cross-reference mismatches between sections. Second, it sometimes produced overly cautious findings. They named the failure first: the evaluator was not independent, because the same agent both reads and judges, so it talks itself out of the mismatches. That points to actor-critic, not a swarm. They added one critic agent with a different system prompt and a different model. They kept the original agent as the actor and re-measured. F1 rose to 0.87 and latency to 18s, both within budget. They wrote the step up as a decision record. It named the conditions for climbing further, such as cross-jurisdictional documents that need a specialist critic. A junior PM had proposed a five-agent democratic council for the same problem. The methodology produced a two-agent system that hit the quality bar at a fraction of the coordination cost. The decision record later stopped a re-escalation impulse when a new failure appeared. The team revisited it, saw the new failure was not on the recorded triggers, and fixed it inside the existing actor-critic shape instead.

Inputs

Single-agent baseline metrics — Numbers that show where the single agent succeeds and where it fails.
Coordination cost budget — The latency, tokens, and infra you can spend on adding more agents.

Outputs

Agent-count decision — A written choice of one, a few, or many agents, with the evidence that backs it.
Coordination shape — The named coordination style (democratic, manager, hierarchical, actor-critic, or swarm) with clear roles.

Steps (6)

Build and measure the single-agent baseline
Build the simplest single agent that could plausibly solve the task. Measure its success, latency, and failures on a representative eval set.
usesAugmented LLM Tool Use
Identify the specific failure that escalation should fix
Name the failure the second agent will fix: context overflow, role conflict, a missing specialist skill, or a judge that is not independent. Without a named failure, adding agents is just theatre.
Pick the coordination style
Choose on purpose. Democratic: peers vote. Manager: one orchestrator with workers. Hierarchical: supervisors nested over sub-teams. Actor-critic: one agent generates, another judges. Each one fits a different failure.
usesOrchestrator-Workers Supervisor Debate Evaluator-Optimizer Hierarchical Agents
Add the smallest number of agents that resolves the failure
Two is often enough. Do not leap to a swarm. Extra agents add coordination cost and bring new failures of their own.
Re-measure against the same eval set
Compare the multi-agent system to the single-agent baseline on the same metrics. If the gains are small, roll back.
Document the escalation as a decision record
Write down why you stepped up, what failure it fixed, and when stepping back down would be right. The next team needs this.

Framework-specific instructions

Pick a framework and generate a framework-targeted rewrite of this methodology's steps.

Choose framework

AI-generated for Agent Development Kit (ADK) (Google) — verify against official docs.

Principles

Single agent first. Multi-agent is a tuning move, not a starting point.
Every step up is paired with a named failure it fixes.
Coordination style is a deliberate pick from a short list, not the framework default.
Re-measure after stepping up. Roll back if the gains are small.

Agent Count Escalation

Methodology process overview

Steps (6)

Build and measure the single-agent baseline

Identify the specific failure that escalation should fix

Pick the coordination style

Add the smallest number of agents that resolves the failure

Re-measure against the same eval set

Document the escalation as a decision record

Framework-specific instructions

Principles

Known failure modes (4)

Related patterns (6)

Related compositions (2)

Related methodologies (2)

Sources (2)

Provenance