{"slug":"multi-agent-systems","label":"Multi-Agent Systems","item_count":3,"day_count":3,"source_count":2,"first_seen":"2026-09-18T09:25:06+00:00","last_updated":"2026-09-30T14:09:57+00:00","generated_at":"2026-10-01T05:03:30.930161+00:00","sources":["arxiv_agent_systems_research","arxiv_cs_lg"],"days":[{"date":"2026-09-18","items":[{"title":"MACE: Memory-Agent Co-Evolution with Adaptive Memory Graphs for Multi-Agent Systems","url":"http://arxiv.org/abs/2609.21533v1","source":"arxiv_cs_lg","type":"paper","summary_1line":"LLM-based multi-agent systems generate collaboration traces that record how agents plan tasks, verify intermediate results, and repair failures. Reusing these procedures requires preserving an action's prerequisites a...","why_it_matters":"Matches feed focus: agent, eval.","sid":"9463df6fcb6ce102","published":"2026-09-18T09:25:06+00:00","editor_note":"Memory-graph approach to reusing multi-agent collaboration traces."}]},{"date":"2026-09-29","items":[{"title":"Topological Coherence for Self-evolving Multi-agent Systems","url":"http://arxiv.org/abs/2609.37953v1","source":"arxiv_agent_systems_research","type":"paper","summary_1line":"Complex tasks inherently couple workflow structure, agent responsibility, collaboration, and memory access: task regions delimit responsibility and tool scope, cross-region dependencies give rise to handoffs, and owne...","why_it_matters":"Matches feed focus: agent, eval.","sid":"c9bd4843a93b6790","published":"2026-09-29T16:27:14+00:00","editor_note":"Frames handoffs and memory access as properties of workflow structure."}]},{"date":"2026-09-30","items":[{"title":"Safety of Latent Communication in Multi-Agent Systems","url":"http://arxiv.org/abs/2609.39788v1","source":"arxiv_agent_systems_research","type":"paper","summary_1line":"Latent communication enables multi-agent systems to exchange information directly in internal representation space, reducing the token, computation, and latency overhead of text-based communication. To this end, light...","why_it_matters":"Matches feed focus: agent, eval.","sid":"9b401e98b8571b23","published":"2026-09-30T14:09:57+00:00","editor_note":"Raises safety concerns for agents that communicate in representation space."}]}],"editorial":{"tldr":"Three arXiv papers in two weeks treat multi-agent systems as a shared-state problem. MACE (Sep 18) stores collaboration traces in an adaptive memory graph; a Sep 29 paper ties task regions to agent responsibility, handoffs, and memory access.","stale":false,"whats_new":"A Sep 30 paper examines safety of latent communication, where agents exchange internal representations instead of text to cut token and latency cost.","why_it_matters":"Memory, handoffs, and inter-agent channels are where multi-agent systems break in production; these papers propose mechanisms, not production results.","take_for_builders":"Keep inter-agent messages as inspectable text in anything audited; treat latent communication as research until a safety evaluation replicates.","status":{"state":"Research · developing","tone":"rising","changed":"2026-09-30","detail":"Three papers, all single-paper claims with no independent replication yet."},"beats":[{"kicker":"MEMORY","tone":"launch","headline":"MACE reuses collaboration traces through adaptive memory graphs","summary":"Preserves each action's prerequisites so agents can reuse plan, verify, and repair procedures.","sids":["9463df6fcb6ce102"]},{"kicker":"STRUCTURE","tone":"rising","headline":"Topological coherence couples workflow, responsibility, and memory access","summary":"Task regions define tool scope; cross-region dependencies become handoffs.","sids":["c9bd4843a93b6790"]},{"kicker":"NOW","tone":"now","headline":"Latent communication trades text overhead for a new safety surface","sids":["9b401e98b8571b23"]}],"open_questions":["Does any of these approaches ship in an open-source agent framework?","Do latent channels get an independent safety evaluation beyond the proposing paper?"],"generated_at":"2026-10-01T05:10:00+00:00"}}