Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarAWS Machine Learning Blog→ original

AWS: A Chain of 10 Agents at 95% Reliability Each Succeeds Only 59.9% of the Time

Multi-agent systems usually break at the handoffs between agents, not because of the model. If each of 10 steps in a chain succeeds 95% of the time, only 59.9% of runs finish without a failure. At 20 steps, that drops to 35.8% (without retries). AWS therefore advises starting with a fixed, predefined workflow and letting the model choose the path only when the steps can't be described in advance.

AI-processed from AWS Machine Learning Blog; edited by Hamidun News
AWS: A Chain of 10 Agents at 95% Reliability Each Succeeds Only 59.9% of the Time
Source: AWS Machine Learning Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceAWS Machine Learning Blog ↗
Editor-in-Chief
Key metric #1A god agent with 15+ tools in one prompt gets confused when selecting them; the limit is five
Unlimited retries burn tokensthe stop condition must live in code
Key metric #3Agent debates are the most expensive pattern, reserved for rare, high-cost decisions

Picture a pipeline of agents: the first extracts data from a document, the second transforms it, the third validates it, and the fourth sends the result onward. The demo works, but in production the system keeps spitting out garbage. In its write-up, AWS explains that the culprit is usually not the model but coordination.

This is a classic distributed-systems problem, and many teams taking it on don't yet have the engineering habits it requires. Hence the headline figure: if each of 10 steps in a chain succeeds 95% of the time, only 59.9% of runs get through the whole path without a failure, and with 20 steps, only 35.

8%. The calculation is simplified, since it assumes a strictly sequential chain with no retries, but the direction is clear: every extra handoff between agents multiplies risk.

The practical takeaway: before adding agents, work out whether you need them at all. According to data AWS cites, about 88% of AI pilots never reach production. A Stanford HAI study found that two models working together can perform worse than either one alone if the cost of coordinating them outweighs the benefit.

Two Levels of Patterns: Who Chooses the Path

AWS classifies coordination schemes by who controls the flow of work. At the first level, the developer defines the path. At the second, the model chooses it at runtime. The second option is more flexible, but it costs more in tokens, runs slower, and is harder to debug. The working rule: start at the first level and move a task to the second only if its subtasks emerge as the work progresses.

First level (fixed workflow):

  • Sequential pipeline: A → B → C. The order is known in advance, so it is easy to test and the cost is predictable. It suits document processing, ETL, and warranty claim handling.
  • Conditional workflow: after a check, the work follows a branch. For example, the code is written, the tests fail, and the task goes back to the author no more than N times before escalating to a human. The orchestrator enforces the limit, for example through an iteration cap in AWS Step Functions.
  • Router: a classifier reads the request and sends it to exactly one specialist, such as quality, maintenance, or scheduling.
  • Parallel workflow: the task is split into independent parts, and the results are merged afterward. Latency equals that of the slowest agent, not the sum of all of them. If one branch fails, only that branch is retried.

Second level (the model chooses the path):

  • Plan-based orchestration: a planner agent maintains an explicit task list, while stateless executors see only their own fragment. If one fails, the others are unaffected.
  • Handoff: an agent passes a task to a colleague together with verified state and evidence. If only a chat summary is passed along, it is not a handoff.
  • Agent conversation: a shared chat in which turn order is set round-robin or by the model, and state is stored in the conversation history.
  • Debate and consensus: several agents independently evaluate the same input, then a majority or a judge decides. This is the most expensive pattern: on perception tasks it buys latency, not accuracy. Keep it for rare decisions where a mistake costs more than the deliberation.

In practice, most production systems are hybrid: fixed workflows for the well-understood parts and autonomous agents where flexibility is needed.

Five Anti-Patterns That Burn Budget

  1. God agent. One agent handles equipment effectiveness analysis, controller diagnostics, and maintenance, with 15 or more tools in a single prompt. Tool selection gets worse, and errors are hard to attribute. It is better to use several agents with at most five tools each.
  2. Unlimited retries. The agent loops on "try again" with no hard stop. No single call raises an alarm, yet tokens keep draining. An iteration ceiling, a per-agent token limit, and a budget signal with automatic shutdown must live in code. The model will not stop on its own.
  3. Natural-language contracts. Agents pass text to each other with no schema. "Possible wear" turns into "confirmed wear" and then into a repair order. The drift happens between agents, where tracing is thinnest.
  4. Premature multi-agent design. Each additional agent adds a nonlinear coordination tax, and the benefit plateaus at around four. The test: if one agent copes simply by switching tool groups, it is not a multi-agent task.
  5. An agent with no protection at the customer boundary. Emails, service visits, and machine parameters are triggered without confirmation. Hallucinations spill into the physical world, and liability escalates from refunds to product liability. AWS proposes layered protection: recommendation, Amazon Bedrock Guardrails, schema validation, business rules, human approval, and only then execution.

What to Take Away

Choose the simplest architecture that meets the requirements. If you can draw the flow on a whiteboard before launch, start at the first level. Keep limits, data schemas, and safety rules in infrastructure and code, not in prompts. In manufacturing, takt time, safety boundaries, edge data, and shift cycles come on top of this, and they determine which pattern fits.

Story

Key Questions Answered

What are the key details on Two Levels of Patterns: Who Chooses the Path?

AWS classifies coordination schemes by who controls the flow of work. At the first level, the developer defines the path. At the second, the model chooses it at runtime. The second option is more flexible, but it costs more in tokens, runs slower, and is harder to debug. The working rule: start at the first level and move a task to the second only if its subtasks emerge as the work progresses. First level (fixed workflow): Sequential pipeline: A → B → C.

What are the key details on Five Anti-Patterns That Burn Budget?

God agent. One agent handles equipment effectiveness analysis, controller diagnostics, and maintenance, with 15 or more tools in a single prompt. Tool selection gets worse, and errors are hard to attribute. It is better to use several agents with at most five tools each. Unlimited retries. The agent loops on "try again" with no hard stop. No single call raises an alarm, yet tokens keep draining. An iteration ceiling, a per-agent token limit, and a budget signal with automatic shutdown must live in code.

What to Take Away?

Choose the simplest architecture that meets the requirements. If you can draw the flow on a whiteboard before launch, start at the first level. Keep limits, data schemas, and safety rules in infrastructure and code, not in prompts. In manufacturing, takt time, safety boundaries, edge data, and shift cycles come on top of this, and they determine which pattern fits.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Autonomous AI Execution · Nomad Agent

Delegate Deep Research, Coding & Multi-Step Workflows to Nomad Agent

Deploy autonomous task execution with Nomad Agent (agent.nomadnet.ai) — automated web research, data synthesis, code generation, and MCP tool orchestration.

Launch Nomad Agent →Nomad AI Workspace
What do you think?
Loading comments…