Agentic AI Multi-Agent Architecture Series - Part Two: Deployment at Scale

Continues from Part One. The first paper covered the shift to coordinated systems, why single agents break at scale, the five architecture patterns, infrastructure, self-replication and governance.

Read Part One →

THE SHORT ANSWER

A multi-agent design that works in a pilot tells you almost nothing about how it behaves in production. Agentic workflows consume 5–30x more tokens per task than a standard chatbot, because one user request can trigger ten to twenty model calls. Around 40% of multi-agent pilots fail within six months of deployment, and Gartner expects over 40% of agentic projects to be cancelled by the end of 2027.

The failures are not prompt problems. Research across five frameworks and 150+ tasks found 14 distinct failure modes and concluded most are structural. This paper covers what changes between pilot and production i.e. economics, observability, identity and state, and sets out the five controls and the operating cadence that keep a fleet alive once it meets real traffic.

In December 2025, Uber's engineering organisation was at 32% adoption of an agentic coding tool. By March 2026 it was at 84%. By April, the company's entire annual AI budget was gone, and its CTO told the press he was back at the drawing board because the budget he thought he would need had already been blown away. Monthly API costs were running between five hundred and two thousand dollars per engineer.

That is not a story about a company with weak financial controls. It is a story about a cost curve almost nobody models correctly, because the thing that makes agentic systems valuable is the same thing that makes them expensive: they do not answer, they work. A chatbot query triggers one inference call. An agentic workflow plans, calls tools, checks its own output, notices a problem, and tries again (ten to twenty model calls for a single user-initiated task). Gartner's March 2026 analysis puts the multiplier at five to thirty times the tokens per task of a standard chatbot.

Part One of this series ended at the point most architecture papers end: a designed system, a chosen pattern, a governance model on paper. This paper is about the part that actually decides the outcome i.e. the eighteen months after go-live, when the design meets volume it was never tested against, a bill nobody forecast, and incidents that were theoretical until they weren't.

Deployment is not the last step of architecture. It is a distinct discipline with its own failure modes, and it is where the majority of multi-agent value is currently being lost.

01 · THE ECONOMICS: The Token Multiplier Nobody Modelled

Start with the number that ends most programmes, because it arrives before any of the others. Pilot economics and production economics for the same multi-agent design are not on the same scale. A pilot runs perhaps fifty to five hundred controlled queries with predictable inputs. Production runs tens of thousands of requests daily, with edge cases, retries, concurrency and real business stakes. A three-agent workflow that costs tens of dollars across a demo cycle can generate five-figure monthly bills at production volume.

The multiplier compounds through three mechanisms, and understanding them is what makes the cost forecastable rather than surprising.

First, call multiplication: each step in an agent's loop i.e. plan, act, observe, verify, correct, is a separate inference.

Second, context carriage: in naive designs every agent receives the full accumulated context, so token consumption grows with the product of agent count and context length. Measurements on production five-agent code-review pipelines record 42,000 to 71,000 tokens per invocation, of which 29–38% is redundant context consumed by agents that never act on it.

Third, retry amplification: a failure at step nine of a twelve-step graph may re-run everything upstream of it.

The practical response is not to abandon multi-agent designs; it is to treat cost as an architectural property rather than a billing surprise. That means budgeting for the multiplier before deployment, scoping context per agent instead of broadcasting it, caching intermediate reasoning, routing simple sub-tasks to cheaper models, and instrumenting token consumption per workflow step as a first-class metric alongside latency and accuracy. A fleet without a per-task spend cap and a circuit breaker is a fleet with an unbounded liability attached to it.