Papers on agentic and multi-agent techniques (MAS) skyrocketed from 820 in 2024 to over 2,500 in 2025. This surge means that MAS are actually a major focus for the world’s prime analysis labs and universities. But there’s a disconnect: Whereas analysis is booming, these techniques nonetheless steadily fail once they hit manufacturing. Most groups instinctively attempt to repair these failures with higher prompts. I exploit the time period prompting fallacy to explain the assumption that mannequin and immediate tweaks alone can repair systemic coordination failures. You’ll be able to’t immediate your approach out of a system-level failure. In case your brokers are persistently underperforming, the difficulty doubtless isn’t the wording of the instruction; it’s the structure of the collaboration.
Past the Prompting Fallacy: Frequent Collaboration Patterns
Some coordination patterns stabilize techniques. Others amplify failure. There isn’t a common greatest sample, solely patterns that match the duty and the way in which data must circulation. The next offers a fast orientation to frequent collaboration patterns and once they are likely to work nicely.
Supervisor-based structure
A linear, supervisor-based structure is the commonest place to begin. One central agent plans, delegates work, and decides when the duty is finished. This setup may be efficient for tightly scoped, sequential reasoning issues, akin to monetary evaluation, compliance checks, or step-by-step determination pipelines. The power of this sample is management. The weak point is that each determination turns into a bottleneck. As quickly as duties develop into exploratory or artistic, that very same supervisor typically turns into the purpose of failure. Latency will increase. Context home windows refill. The system begins to overthink easy choices as a result of all the things should cross by a single cognitive bottleneck.
Blackboard-style structure
In artistic settings, a blackboard-style structure with shared reminiscence typically works higher. As an alternative of routing each thought by a supervisor, a number of specialists contribute partial options right into a shared workspace. Different brokers critique, refine, or construct on these contributions. The system improves by accumulation fairly than command. This mirrors how actual artistic groups work: Concepts are externalized, challenged, and iterated on collectively.
Peer-to-peer collaboration
In peer-to-peer collaboration, brokers trade data instantly and not using a central controller. This could work nicely for dynamic duties like net navigation, exploration, or multistep discovery, the place the aim is to cowl floor fairly than converge shortly. The chance is drift. With out some type of aggregation or validation, the system can fragment or loop. In follow, this peer-to-peer fashion typically reveals up as swarms.
Swarms structure
Swarms work nicely in duties like net analysis as a result of the aim is protection, not rapid convergence. A number of brokers discover sources in parallel, comply with completely different leads, and floor findings independently. Redundancy isn’t a bug right here; it’s a characteristic. Overlap helps validate alerts, whereas divergence helps keep away from blind spots. In artistic writing, swarms are additionally efficient. One agent proposes narrative instructions, one other experiments with tone, a 3rd rewrites construction, and a fourth critiques readability. Concepts collide, merge, and evolve. The system behaves much less like a pipeline and extra like a writers’ room.
The important thing threat with swarms is that they generate quantity quicker than they generate choices, which may additionally result in token burn in manufacturing. Contemplate strict exit circumstances to stop exploding prices. Additionally, and not using a later aggregation step, swarms can drift, loop, or overwhelm downstream parts. That’s why they work greatest when paired with a concrete consolidation part, not as a standalone sample.
Contemplating all of this, many manufacturing techniques profit from hybrid patterns. A small variety of quick specialists function in parallel, whereas a slower, extra deliberate agent periodically aggregates outcomes, checks assumptions, and decides whether or not the system ought to proceed or cease. This balances throughput with stability and retains errors from compounding unchecked. This is the reason I train this agents-as-teams mindset all through AI Brokers: The Definitive Information, as a result of most manufacturing failures are coordination issues lengthy earlier than they’re mannequin issues.
When you suppose extra deeply about this group analogy, you shortly notice that artistic groups don’t run like analysis labs. They don’t route each thought by a single supervisor. They iterate, focus on, critique, and converge. Analysis labs, however, don’t function like artistic studios. They prioritize reproducibility, managed assumptions, and tightly scoped evaluation. They profit from construction, not freeform brainstorming loops. This is the reason it’s not a shock in case your techniques fail; for those who apply one default agent topology to each downside, the system can’t carry out at its full potential. Most failures attributed to “unhealthy prompts” are literally mismatches between job, coordination sample, data circulation, and mannequin structure.
Need Radar delivered straight to your inbox? Be part of us on Substack. Join right here.
Breaking the Loop: “Hiring” Your Brokers the Proper Approach
I design AI brokers the identical approach I take into consideration constructing a group. Every agent has a talent profile, strengths, blind spots, and an applicable function. The system solely works when these abilities compound fairly than intrude. A powerful mannequin positioned within the mistaken function behaves like a extremely expert rent assigned to the mistaken job. It doesn’t merely underperform, it actively introduces friction. In my psychological mannequin, I categorize fashions by their architectural character. The next is a high-level overview.
Decoder-only (the mills and planners): These are your customary LLMs like GPT or Claude. They’re your talkers and coders, sturdy at drafting and step-by-step planning. Use them for execution: writing, coding, and producing candidate options.
Encoder-only (the analysts and investigators): Fashions like BERT and its fashionable representations akin to ModernBERT and NeoBERT don’t speak; they perceive. They construct contextual embeddings and are glorious at semantic search, filtering, and relevance scoring. Use them to rank, confirm, and slender the search area earlier than your costly generator even wakes up.
Combination of consultants (the specialists): MoE fashions behave like a set of inside specialist departments, the place a router prompts solely a subset of consultants per token. Use them whenever you want excessive functionality however wish to spend compute selectively.
Reasoning fashions (the thinkers): These are fashions optimized to spend extra compute at check time. They pause, replicate, and verify their very own reasoning. They’re slower, however they typically stop costly downstream errors.
So if you end up writing a 2,000-word immediate to make a quick generator act like a thinker, you’ve made a foul rent. You don’t want a greater immediate; you want a unique structure and higher system-level scaling.
Designing Digital Organizations: The Science of Scaling Agentic Programs
Neural scaling1 is steady and works nicely for fashions. As proven by basic scaling legal guidelines, rising parameter depend, knowledge, and compute tends to end in predictable enhancements in functionality. This logic holds for single fashions. Collaborative scaling,2 as you want in agentic techniques, is completely different. It’s conditional. It grows, plateaus, and generally collapses relying on communication prices, reminiscence constraints, and the way a lot context every agent really sees. Including brokers doesn’t behave like including parameters.
This is the reason topology issues. Chains, timber, and different coordination buildings behave very in another way below load. Some topologies stabilize reasoning as techniques develop. Others amplify noise, latency, and error. These observations align with early work on collaborative scaling in multi-agent techniques, which reveals that efficiency doesn’t enhance monotonically with agent depend.
Current work from Google Analysis and Google DeepMind3 makes this distinction specific. The distinction between a system that improves with each loop and one which falls aside isn’t the variety of brokers or the dimensions of the mannequin. It’s how the system is wired. Because the variety of brokers will increase, so does the coordination tax: Communication overhead grows, latency spikes, and context home windows blow up. As well as, when too many entities try to resolve the identical downside with out clear construction, the system begins to intrude with itself. The coordination construction, the circulation of data, and the topology of decision-making decide whether or not a system amplifies functionality or amplifies error.
The System-Degree Takeaway
In case your multi-agent system is failing, pondering like a mannequin practitioner is not sufficient. Cease reaching for the immediate. The surge in agentic analysis has made one fact plain: The sector is transferring from immediate engineering to organizational techniques. The subsequent time you design your agentic system, ask your self:
- How do I manage the group? (patterns)
- Who do I put in these slots? (hiring/structure)
- Why may this fail at scale? (scaling legal guidelines)
That mentioned, the winners within the agentic period gained’t be these with the neatest directions however the ones who construct probably the most resilient collaboration buildings. Agentic efficiency is an architectural end result, not a prompting downside.
References
- Jared Kaplan et al., “Scaling Legal guidelines for Neural Language Fashions,” (2020): https://arxiv.org/abs/2001.08361.
- Chen Qian et al., “Scaling Giant Language Mannequin-based Multi-Agent Collaboration,” (2025): https://arxiv.org/abs/2406.07155.
- Yubin Kim et al., “In direction of a Science of Scaling Agent Programs,” (2025): https://arxiv.org/abs/2512.08296.