Multi-Agent Architecture Patterns in ProductionMixed sources18 sources, drawn from a mix of official docs, project repositories and research.This rates where the sources come from. It doesn't prove every claim.
Orchestrator, supervisor, and swarm patterns for multi-agent systems with real trade-offs and failure modes.
By Amit UpadhyayPublished Updated 14 minArticle 14 of 16 in Agentic Systems
AI assistant · answers from this article
1 of 5
Jobs that finish in seconds
Nothing to add yet
When a whole job takes seconds, restarting it is the simplest recovery there is.
On this page
Prerequisites
Before reading this article, make sure you're familiar with:
Multi-agent orchestration is what keeps several agents working on one job reliable: saved hand-offs, retries for one worker at a time, limits on time and cost, and one trace across them all.
You reach for it when several agents work on one job in production, and any of them can fail, stall or overspend. Not before: it costs a job store or queue, retry logic, and tracing work across agent boundaries.
The cheat sheet
What to use when
You do not need all of this on day one. Each row is one signal and the piece that answers it. Stop at the row your system has reached, or pick one to see it in the picture above.
Where you areThe signalWhat to useWhat it costs
None. A failed job costs almost nothing to run again.Nothing. Run it in one process, and rerun on errorNothing
One failed worker throws away the finished work of all the others.Save each agent’s result as soon as it finishesA job store or a queue to run
A rate-limit error in one worker restarts the entire job.Retry the failing worker with backoff, not the jobRetry logic for each worker, and a limit on it
One stuck worker holds the job open for hours, and the tokens keep adding up.A deadline and a token budget for each jobIteration limits ↓Some jobs delivered incomplete
A brief is wrong, and you cannot tell which agent introduced the error.One trace ID across every agent in the jobNo observability ↓Tracing carried across agent boundaries
Start With One Agent
One agent with a handful of good tools is the right first design. It's cheaper, and there's one context to read when something goes wrong. For a task that needs one kind of skill, it's also the right last design. Splitting the work is worth considering once that agent holds so many tools and instructions that it does every part half well.
Give one agent research, planning, coding, review and deployment, and you get a very long system prompt, a pile of tools and mediocre results at every step. Instructions compete for attention, the wrong tool gets picked, and the agent loses track of what it already found.
Splitting the work across specialist agents can fix that, but it is not free: more tokens, more hand-offs, and failures that pass from one agent to the next. The patterns below (orchestrator, supervisor and swarm) differ mostly in how they coordinate, and picking the wrong one costs more than staying with a single agent.
Signs you have outgrown one agent
Your single-agent system works for demos but produces mediocre output on complex, multi-step tasks
You've seen the "agent swarm" videos on Twitter and wonder if that's actually how production systems work
You tried chaining multiple LLM calls and the context window exploded after 3 rounds
You want multi-agent but have no idea which pattern fits your use case
Pro deep dive14 min full read
Keep reading with Pro
This is one of our advanced production deep dives. Pro unlocks the rest of it, every other deep dive, and the See, Animate, Test and Build modes on every topic.
What it covers
Orchestrator, supervisor, and swarm patterns for multi-agent systems with real trade-offs and failure modes.