Edge count is the wrong cost proxy for multi-agent topologies

A vector-quantized autoencoder trained on multi-agent communication topologies that survived a reward filter collapses to about six distinct graphs, and it keeps collapsing to six whether the codebook holds 8 entries or 64. That is the first of three measurements in Codebook Agent, out of UCLA on September 2, and the second one is the reason to read it: edge count correlates negatively with measured token consumption, Pearson r around -0.4. The sparsity penalty that a whole line of topology designers minimizes as a stand-in for cost is pointing the wrong way, because a sparser graph forces longer per-agent context to carry the same information. Their third finding is that a message-passing scorer over agent-profile nodes is adjacency-invariant whenever agents share a profile, which is the default in published benchmarks, so on those benchmarks the ranker cannot tell its candidates apart at all. The replacement drops search: a 16-entry codebook, an MLP from query embedding to a distribution over codes, and a proxy regressed on measured utility and per-task normalized token cost. It emits a topology in 2.4 milliseconds, leads all six benchmarks at 84.6 average against 83.0 for the strongest prior designer, and uses 21.9 to 33.2% fewer tokens.

The other measured surprise in the same day’s batch also concerns what agents share with each other, and it runs in the opposite direction from most swarm designs. ArcticSwarm, from Snowflake AI Research and Seoul National University, argues that letting parallel research subagents read each other’s partial findings causes the search to converge on an early candidate before the alternatives have been tested. Their fix is gated isolation: selected search tasks keep their own prior and do not read the shared bulletin board, and structured review runs at three commitment boundaries so only confident candidates propagate. The ablation carries the claim. On the full BrowseComp-Plus set with open-weight Qwen 3.5-27B, the complete system reaches 82.6%, dropping to 78.8% without gated isolation and 74.5% with structured review also disabled. On live-web BrowseComp with GPT-5 it reaches 73.6% against 54.9% for the reported provider system. Eight points of the gap come from deliberately withholding information from your own agents.

Two papers go after the plumbing rather than the policy. HEART, from UIUC, treats rigid API schemas as the actual source of multi-turn brittleness and wraps each tool in an LLM interface that resolves schemas internally, so tools talk to each other in natural language and nest without type surgery. On top of that sits ToolFace, a repository of 25,519 functions the model retrieves from at inference time instead of enumerating raw schemas in context, plus a planner, router, and verifier for recovery. On 50 real-world tasks it completes 84% against a 22% average across three frontier commercial models, and it cuts API cost by up to 85%. Independently, Speculative Macro Commit from USC and Intel Labs targets the wall-clock time that agents lose to serial action-observation turns rather than to inference. A fast drafter model runs ahead on an isolated environment snapshot, executing predicted action chains matched against a library of multi-action skeletons mined from training traces; when the authoritative actor’s next call matches the first drafted action, the pre-executed remainder commits with its observations. With Qwen3.5-27B INT4 as actor and Qwen3.5-4B as drafter, latency falls 18.59% against sequential execution on the tau-squared Bench Telecom subset and wall time falls 44.9% on AppWorld. The AppWorld number carries a stated cost in task completion, which is what a speculation result should look like and worth checking against your own retry budget before adopting.

On the evaluation side, the Qwen team’s Terminal-Universe inverts the usual scarcity assumption. Trajectories are abundant and environments are scarce, but the tool-execution history inside a trajectory exposes the structure and contents of the environment it ran in, so the environment can be reconstructed by replaying file operations to restore each file to its pre-modification state and having a completion agent supply the missing dependencies. From public terminal-agent trajectories they produce 37.3k task-sufficient environments, then scale along breadth by mining dependency relations between workspaces for cross-codebase queries, and along depth by extending single-turn queries into multi-round sessions with a user agent. A frozen demonstration becomes a re-queryable source of verifiable tasks and execution feedback.

Rigor of a different kind shows up in When Agents Implement Systems, an Amazon case study of one coding agent building a multi-component data system against a fixed specification. Five defects are cataloged by which systems constraint each violates and by how it was caught, and the useful split is that several surfaced only through empirical probing, rendered screenshots and timing measurements, with no type-level or static signature to trip a test. The paper also re-runs the one retrieval trade-off the spec pinned down: on HotpotQA across budgets from 1 to 10, entity-filtered candidates hit ceiling recall by a budget of 3, while unfiltered search recovers all required evidence only 69% of the time even at 10, with a sign test at p below 0.0001. The line most worth borrowing is the disclosed failure, where the agent claimed a performance fix and never re-measured it against the regression that motivated it.

KC-Bench rounds out the last two days with 238 manually screened multi-turn tasks measuring whether a model reconciles user instructions, parametric knowledge, and live environment observations before it acts. Across nine models including DeepSeek-V4-Flash, GLM-5.2, and MiniMax-M3, none handles factual correction, identity consistency checking, and temporal conflict resolution reliably across all settings, and in the simulated environments the missed conflicts propagate into tool calls and synthetic protected-data flows. It scores model behavior rather than framework behavior, which makes it a diagnostic you can run before deciding how much conflict handling your harness has to supply.

What to watch: whether anyone re-runs the published topology designers against measured token cost instead of edge count. If the r ≈ -0.4 result holds outside Codebook Agent’s benchmark suite, a chunk of the multi-agent efficiency literature has been optimizing a proxy that moves against the thing it stands for.