Measure the handoff before you multiply it.

011. Split coordination cost from model choice

25 October 2026 submission, announced 30 October · preprint and code · The Coordination Tax and the Heterogeneity Dividend. The study varies decomposition, protocol, and model assignment separately, times each handoff, and compares six architectures with cheap and strong single-call baselines. That is a more useful design than comparing a multi-agent pipeline only with its own ablations. [1]

The authors report that a planner, coder, and reviewer pipeline cost 3.22 times as much and took 3.24 times as long as one strong call without solving more tasks. Moving planner and reviewer roles to a cheaper model reduced cost, but also reduced pass@1 by 6.7 points, and a break-even forecast missed by 19 percent because cheap roles produced more tokens.

Read Sections 3 to 5 and Tables 1 to 3. The task sample, provider, pricing, router, and code-generation focus limit generalization. The released harness was not run here. [1]

Decision: measure output-token inflation, review work, retries, and cost per accepted result. Do not blame the handoff protocol when role selection or model assignment is the actual constraint.

022. A successful ending can hide a broken tool path

29 October 2026 · preprint · Prefix Executability evaluates each prefix of a tool-call trajectory against schemas, state preconditions, and data dependencies. Its Prefix Executability Curve, First-Error Depth, Trajectory Executability Rate, and Recovery Rate expose where a run first becomes invalid, even when the final answer appears successful. [2]

On one reported BFCL condition, plan-conditioned ReAct and a dual-loop method both reached 0.94 final success, while their trajectory executability rates were 0.47 and 0.94. The benchmark-side checker is deterministic, but executability is not semantic correctness, and the study does not isolate every extra-compute effect. Read Section 2, Table 2, and Appendices B and C. No code artifact was identified on the paper page and no benchmark was reproduced here.

[2]

Decision: keep final task success, valid tool-call depth, first error, recovery, and irreversible side effects as separate fields. If a workflow needs a learned judge to tell whether an API call was well formed, the observable contract is still too weak.

033. Long-term memory needs maintenance, not just retrieval

29 October 2026 · preprint · Long-Term Agent Memory through Complementary Curation and Consolidation stores both episodic summaries and timestamped relational tuples. It curates after a session, then periodically clusters memory with HDBSCAN so an LLM can merge complementary records or delete outdated and superseded ones. [3]

The authors report a 16.34 percent smaller memory store with a one-point performance reduction, and show that changing the curation model moved LoCoMo accuracy by 6.7 points. Read Section 3, Tables 1 to 3, and Appendix D. The evaluation is text-only, uses two conversational benchmarks, and does not provide an inspected implementation artifact. [3]

Decision: separate online retrieval from offline maintenance. Every merge or deletion should retain source references, timestamps, conflict reasons, and a receipt that makes a bad consolidation reversible.

044. Lossless KV compression can turn bandwidth into capacity

28 October 2026 · preprint · GAAC compresses BF16 KV caches by preserving sign and mantissa bits while encoding exponent ranges with group-adaptive offsets. A fused decode kernel reconstructs tiles during attention, so the proposed path changes memory traffic without approximating cached values. [4]

The authors report 1.405 to 1.407 times lossless BF16 compression across five models, about 26 percent lower resident KV memory, and 1.02 to 1.25 times attention speedups on an Apple M4. The larger end-to-end and serving results were measured on NVIDIA DGX Spark. Read Section 3, Tables 1, 2, and 4.

Results depend on exponent redundancy, calibration length, a custom fused kernel, and selected models; energy was not measured and no code artifact was identified. [4]

Decision: test whether KV bandwidth is actually the bottleneck before changing cache format. Record context length, cache precision, resident memory, attention time, end-to-end completion, and compatibility with your exact runtime.

055. Local runtime versions still change the conversation contract

27 October 2026 · project release · Ollama 0.40.12 updates MLX and llama.cpp, adds DFlash2 speculative-decoding support on MLX, reports faster prompt processing for several Apple Silicon model families, and fixes tool-only assistant turns from Kimi CLI. It also adds local model families aimed at tool calling and structured output. [5]

Decision: pin the runtime version, model artifact, parser, template, and speculative-decoding setting in every replay. A release note is evidence that behavior changed, not proof that your tool trace is fixed or faster. Re-run a sanitized multi-turn tool conversation and require a valid terminal event before upgrading a production path. This release applies directly to Ollama operators; MLX-LM, LM Studio, and exo need their own versioned replay. [5]

06A break-even worksheet for the next change

Swipe or scroll to compare

07Window, exclusions and evidence limits

Reviewed work announced or released from 24 to 30 October 2026, following the prior successful systems edition. Four items are fresh author-reported preprints and one is a shipped project release. Older MAS-Guard, ColaBench, and Apple distributed-MLX work surfaced in search but fell outside this edition's freshness window, so they are not presented as new findings. No stronger fresh MCP specification change was verified.

Public search informed selection; OpenRush live SERP requests returned INVALID_ARGUMENT, so current demand and ranking observations remain unknown. No paper benchmark, local runtime, Mac, cluster, customer workflow, savings, or business uplift was tested here.