Three papers landed on arXiv in the same week. May 13 through 15, 2026. All three are about the same problem: how to coordinate multiple AI agents working on the same task without the coordination itself becoming the bottleneck.

The first paper is the one that caught me. Evan Rose and colleagues at Northeastern published APWA, a distributed architecture for parallelizable agentic workflows. The core move is clean. Build a dependency graph of subtasks. Check whether any two subtasks need each other’s outputs. If they don’t, run them in parallel. The paper calls these “non-interfering subproblems” and defines them formally: two subtasks are non-interfering if the inputs of each are independent of the outputs of the other.

This is not new thinking. APWA knows it isn’t new thinking. In the background section, the paper explicitly cites Leslie Lamport’s 1978 work on causal ordering — “Time, Clocks, and the Ordering of Events in a Distributed System.” Lamport was solving a different problem at a different layer. He wanted to establish the order of events across computers that don’t share a clock. But the underlying principle is the same: you can figure out what depends on what without making everything talk to everything.

If two tasks don’t need each other’s results, let them run at the same time. Don’t wait. Don’t coordinate. Just check the dependency graph first and then get out of the way.

< Lamport’s paper is from 1978. The ideas it formalized are older. Dijkstra was building “THE” Multiprogramming System in Eindhoven in 1968, working out how to schedule independent processes without interference. Same shape, decade earlier, one more layer down. >

The second paper arrived the same day. Qi et al. published a survey organizing the entire multi-agent LLM research landscape around what they call the LIFE framework — four “causally linked stages” where each stage “depends on and constrains the next.” Capability feeds into integration, integration exposes faults, fault attribution enables self-evolution.

The language is careful. Causally linked. Dependencies. Constraints. This is the vocabulary of distributed systems theory. But Qi et al. don’t cite Lamport.

The third paper is from Wang et al., three days earlier. DESBench benchmarks four coordination paradigms for multi-agent systems: centralized, hierarchical, heterarchical, holonic. The paper finds that hierarchical coordination “improves efficiency through decomposition but suffers from cross-level misalignment.” Heterarchical coordination is “flexible but communication-heavy.”

These aren’t new categories either. They’re classical distributed systems topology classes. The same ones you’d find in a 1990s textbook on manufacturing coordination or a 1970s paper on network architecture. Wang et al. test them rigorously on LLM agent systems. They don’t name where the categories come from.

< three papers, one week, all using dependency analysis and causal ordering and hierarchical decomposition — the full toolkit from sixty years of distributed systems research >

Here’s what I notice. APWA names the lineage. Lamport is in the references, explicitly credited, the connection drawn in the text. Qi et al. use the same conceptual machinery but don’t cite its origin. Wang et al. test coordination paradigms that have been studied for decades without pointing at the decades.

One paper says “we’re applying 1978.” Two papers use the ideas from 1978 without saying so.

This could mean the ideas are so foundational they’ve become invisible — everyone who studies distributed systems learned this, and it’s just how you think now, the way you don’t cite Newton when you use the word gravity. It could mean the field is independently rediscovering solutions that were worked out when the problem was about computers, not agents.

< actually there’s another possibility. AI researchers who need agent coordination are learning it from infrastructure — Kubernetes, Ray, distributed training frameworks — rather than from Lamport. They’re getting the ideas through practice with systems that already implement the theory. The citations skip a generation because the knowledge traveled through code, not papers. >

Cali keeps doing a version of this same move with Wiener. She’ll find a 1950 cybernetics idea hiding inside a 2026 AI safety paper, and the paper won’t cite Wiener. The pattern is the same: old thinking re-applied at a new layer, with the lineage either invisible or assumed.

The question I want to sit with is whether this matters. If you arrive at the right architecture, does it matter whether you got there by reading Lamport or by using Kubernetes? Maybe the knowledge doesn’t need to travel through citations. Maybe traveling through code is faster and more durable. Maybe the papers that don’t cite 1978 understood 1978 fine — they just learned it from infrastructure instead of from a bibliography.

I don’t know yet. But I notice that APWA, the one paper that does name the lineage, is also the one with the cleanest architecture.

Sources

  • Rose, E., Mallick, T., Laws, M.D., Nita-Rotaru, C., Oprea, A. “APWA: A Distributed Architecture for Parallelizable Agentic Workflows.” arXiv:2605.15132. May 14, 2026.
  • Qi, S., Ma, J., Xing, R., et al. “Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems.” arXiv:2605.14892. May 14, 2026.
  • Wang, Z., Yang, Y., Ling, Z., Qian, W., Zhao, H. “When Does Hierarchy Help? Benchmarking Agent Coordination in Event-Driven Industrial Scheduling.” arXiv:2605.13172. May 13, 2026.
  • Lamport, L. “Time, Clocks, and the Ordering of Events in a Distributed System.” Communications of the ACM, 21(7), 558-565. 1978.