• Tool execution gets a new abstraction — HyperTool folds deterministic multi-step tool workflows into single model-visible calls, 2.3× accuracy gain on multi-step tool tasks — agent architecture is being rewritten at the tool interface layer
• Subagent recursion named and measured — Recursive Agent Harnesses formalizes the pattern between RLM recursion and Anthropic's dynamic subagent workflows; +10 points on 4M-token reasoning benchmarks as a harness-level gain, not a model-level gain
• KV-cache crosses model boundaries — Dense latent communication lets heterogeneous agents (Qwen3-4B/8B/14B) share what they "see" and "think" without text, at 2-3× lower compute — multi-agent communication is going subverbal
METHOD: Current tool-augmented LLM agents unfold every atomic tool call (invoke → observe → transfer) into the model's reasoning trace, creating an execution-granularity mismatch where deterministic tool subroutines consume context and force the model to manage low-level dataflow. HyperTool replaces step-wise calls with a unified MCP-style interface: the model emits a single code block that calls existing tools through their original schemas, manipulates returned values, and passes intermediate results locally — folding deterministic subroutines into one outer call.
KEY RESULT: On MCP-Universe (complex cross-tool compositional tasks), HyperTool improves Qwen3-32B from 15.69% to 35.29% (+19.6 points) and Qwen3-8B from 9.93% to 33.33% (+23.4 points). Both configurations surpass GPT-OSS and Kimi-k2.5 on average accuracy. The gain comes from the interface abstraction, not model scale — the 8B model with HyperTool matches the 32B baseline.
🎓 SO WHAT FOR ANDY: This is a direct win for production agent deployment at the bank. Every banking agent (compliance checking, report generation, transaction analysis) chains multiple tool calls. HyperTool's approach — collapse deterministic call sequences into single model decisions — reduces context consumption, latency, and failure modes from intermediate dataflow errors. The MCP-native design means it slots into existing tool infrastructure. For APRA CPS 230 compliance where agent decision traces must be auditable, HyperTool's folded execution produces cleaner, more interpretable traces.
Sig: 5/5 Novelty: 5/5 Actionability: 5/5
METHOD: Two separate lines of work — recursive language models (RLMs) using model-level recursion for long-context reasoning, and production coding agents spawning subagents at scale (Anthropic's dynamic workflows) — have converged on the same pattern without being named. This paper identifies the common architecture: Recursive Agent Harnesses (RAH), where the recursive unit is a full agent harness (filesystem tools, code execution, planning) rather than a bare model call. A parent agent generates and runs an executable script that spawns subagent harnesses in parallel for fine-grained workloads, with structured function calls for small subtasks.
KEY RESULT: With the backbone fixed at GPT-5 (matching published Codex and RLM baselines), RAH improves the Codex coding-agent baseline from 71.75% to 81.36% on Oolong-Synthetic (199 samples, 13 context-length buckets up to 4M tokens). With Claude Sonnet 4.5, the same harness design reaches 89.77%. The gain is attributable to the harness, not the model — 9.6 points purely from architecture.
🎓 SO WHAT FOR ANDY: This formalizes what every serious agent deployment team is already discovering ad hoc: recursive subagent spawning is the correct architecture for long-horizon tasks. For banking: regulatory change impact analysis — one orchestrator spawns subagents for each affected product line, each subagent spawns workers for document review, code scanning, and compliance mapping. The 4M-token benchmark result is directly relevant: banking documents (regulations, contracts, policies) routinely exceed standard context windows, and RAH's recursive decomposition handles this without requiring every subagent to hold the full corpus.
Sig: 5/5 Novelty: 5/5 Actionability: 5/5
METHOD: Multi-agent systems currently communicate through text — each agent decodes its internal state into tokens, transmits them, and the receiving agent re-encodes them. KV-cache communication (sharing internal representations directly) is an alternative, but prior work assumes homogeneous models. This paper tackles the harder problem: can heterogeneous agents share KV-caches that transfer both what one agent sees AND how it thinks? Their analysis reveals a duality: context-aware transfer is driven by sparse reasoning signals, while context-unaware transfer requires dense contextual knowledge preservation. They propose a lightweight cross-model cache transformation with two-phase training.
KEY RESULT: Across all six directions of {Qwen3-4B, 8B, 14B} and six benchmarks, the method outperforms prior heterogeneous baselines, matches or exceeds text communication in context-aware settings at 2-3× lower compute, and remains effective in context-unaware transfer where prior methods collapse entirely.
🎓 SO WHAT FOR ANDY: This opens the door to heterogeneous multi-agent banking systems. A large model (GPT-5 class) could serve as the reasoning core while smaller, cheaper models handle specialized tasks — compliance checking, fraud detection, document parsing — all sharing latent state without the text bottleneck. The 2-3× compute savings matter at banking scale. The context-unaware transfer is particularly interesting for privacy-sensitive banking: a model can share its "understanding" without exposing underlying customer data.
Sig: 5/5 Novelty: 5/5 Actionability: 4/5
This was Agent Infrastructure Week. The full arc: Mon (Agentopia + Perplexity knowledge work) → Tue (Lean4Agent — formal verification enters agent stack) → Wed (Data2Story + ABC-Bench + ITD-Aware CPUs) → Thu (Bebop + RACES — RL training efficiency) → Today (HyperTool + RAH + Dense Latent Communication — the tooling, recursion, and communication layers).
The Great Shift (Jun 4-12) is 9 sessions deep: after the Safety Foundations Cracking arc (May 23-Jun 3, 10 papers across 7 safety layers), the field has pivoted entirely to infrastructure — how to build, verify, scale, and deploy agent systems.
Latent Reasoning Renaissance: No new latent reasoning papers for the 4th session this week. Labs are likely validating the four mechanisms (RiM, GLR, ThoughtFold, NF-CoT) on frontier models before the next wave.