The agentic AI field is confidently reinventing distributed systems. Badly.

Here are five sins of your current or future AI system. You will commit all of them. The only choice you’ll make is the order you solve them or the order they would bring your system down.

SIN 1: Coordination is not orchestration.

A central router telling 50+ agents what to do is not a swarm — it’s a monolith with a marketing problem. One biased node, one failure, and your entire “multi-agent system” fails together, in unison, expensively.

Introduce yourself to leaderless consensus algorithm (EPaxos, Raft, etc). DecentLLMs implements it for LLM agents today, tolerating up to 50% Byzantine participants in cooperative tasks. The theory is 40 years old. Adoption in agentic AI is near zero.

Are your agents still phoning home for guidance? I have bad news to you.

SIN 2: You gave your agents stop conditions. Instead give them lifespans.

Every recursion should have a stop condition. Or should it?

We built agentic systems like good engineers: token budgets, recursion limits, no fork-bombs, exit criteria - one theme: responsible… and accidentally optimized for termination instead of renewal.

Nature didn’t give cells a stop condition. It gave them a mortality budget - enforced generational turnover. A healthy cell still dies after enough replications. Not because it failed. Because the system needs the slot. Your best-performing agents are immortal. They accumulate bias, crowd out novelty, converge your system toward a local optimum that feels like it’s working - which is the most dangerous kind.

A stop condition asks: are we done? A lifespan asks: is it time to make room?

Not the same question. Structural mortality is an architectural primitive, not a performance consequence. Design it in.

SIN 3: You are training individuals. You should be training systems.

Tool-wiring 100 agents is not multi-agent learning. It’s single-agent learning, just with a bigger AWS/Azure/tokens bill.

The real problem is joint value optimization - teaching a population to internalize a shared goal so cooperation emerges without being scripted. (MARL research brings the foundations of this, there are more). Infinite-horizon game theory tells you why it works: agents that value the game continuing will always out-cooperate agents chasing local optima.

The goal isn’t agents that follow instructions toward a finish line. It’s the agents that have learned the finish line is the wrong thing to optimize for.

SIN 4: You benchmark for winning. You should benchmark across time.

The Sort Benchmark, 30 years old one - still the gold standard in systems performance, doesn’t just measure who’s fastest. It has separate categories for elapsed time, energy per terabyte, cost per record. Same outcome, multiple cost dimensions, simultaneously.

Your agentic evals measure task completion on finite problems. But the game you’re deploying into is not finite. An agent that wins tasks is not the same as an agent that plays well indefinitely. You have no evals for the latter. At every training run you are selecting for exactly the wrong behavior at scale.

Measure cooperation persistence, strategy stability, behavior under agent turnover. Until you do, your alignment claim is a test result, not a system property.

SIN 5: You cannot claim the emergence you did not design for.

Multi-agent systems are built for emergent behavior. That’s the point. That’s the power.

But “it emerged” is not an engineering claim. It is an observation.

If you didn’t specify the target emergent property — formally, measurably, as a design constraint — you didn’t design for it. You discovered it. Fine for research. Not fine for production. Design for emergence. Instrument for it. Be careful to not to confuse emergence with intent. The gap between the two is where your liability grows uncontrollably.

The theory is decades old. The primitives have existence proof. The gap is epistemic and the most embarrassing one.

Stop building agentic frameworks. Start building distributed cooperative systems.

#MultiAgentSystems #AgenticAI #DistributedSystems #OpenEndedLearning #SystemsEngineering