# Before you add another agent, find the bottleneck.

Five changes worth checking before you buy more agent capacity: stage-specific delegation, bounded shared memory, associative retrieval, MCP placement, and local tool-history fixes.

Published: 2026-10-16 | Updated: 2026-10-16
Author: /init editorial
Method: AI-assisted, source-checked
Canonical: https://init.news/notes/agent-systems-brief-2026-10-16/

## The answer

Spend the next agent experiment on a measured bottleneck. This week’s research separates useful delegation from extra coordination, tests a shared message store with bounded reads, and offers a memory-retrieval alternative to explicit graph construction. Meanwhile, MCP deployment options and local tool-history fixes change integration decisions. Start with one comparison or a conversation replay; none of these sources proves that a larger team will improve your business.

## 1. Delegate the bottleneck, not every stage

12 October 2026 · preprint · Scaling Inference-Time Compute with Teams of Coding Agents. The study varies delegation stage, worker capability and team size instead of treating “multi-agent” as one configuration. Planning-only delegation matched broader delegation at roughly half the inference cost on ProgramBench. But implementation delegation worked best on Commit0-Lite, whose repository skeleton already supplied much of the plan. The paper does not match budgets across configurations and reports empirical operating points, not a guaranteed optimum. [1]

Decision: compare delegation stages on your own task mix before expanding the team. Read Sections 2 and 5–7, then Appendix C.13 for the counterexample; Appendix A describes scheduling and delivery. This is author-reported benchmark evidence, not a reproduced result or a forecast of business savings. [1]

## 2. Repair tool-history compatibility before blaming the model

14–15 October 2026 · releases and code changes · Ollama 0.40.5 and MLX-LM 0.32.0. Ollama fixes continuation of Responses conversations containing custom tool calls. Its patch normalizes custom calls and outputs while checking required fields. Read openai/responses.go and TestResponsesCustomToolHistory in PR #18911. [5] [6]

MLX-LM stops blindly JSON-decoding arguments already supplied as objects; its Qwen3 Coder parser also handles object or array values when a parameter lacks a top-level type. Read process_message_content in PR #1904 and test_qwen3_coder_untyped_param in PR #1910. [7] [8] [9]

Decision: replay a real, sanitized multi-turn tool transcript through the candidate runtime before changing your model or orchestration. These are specific compatibility fixes, not proof of universal API parity or faster inference. Upstream tests were inspected, not executed here.

## 3. Bound what each agent reads from the group

12 October 2026 · preprint · ScaleMAS. Instead of routing everything through a coordinator or broadcasting growing context, agents publish linked work into an append-only datastore, search a bounded set of messages, and visit references. Search combines lexical relevance with citation weight. The authors report improvements across four benchmarks and test larger groups and agent dropout; the 128-agent experiments cover selected tasks, not every benchmark. [2]

Decision: if repeated reading dominates your runs, compare bounded shared retrieval with the current handoff method. Citation popularity is a retrieval signal, not independent truth verification. Read Sections 2.1–2.3 for the storage contract and operations, Section 3.5 for ablations, and Appendix C for implementation settings. The evidence is inference-time research, not a production reliability guarantee. [2]

## 4. Test a second retrieval path before committing to a graph

14 October 2026 · preprint · MemCode. Dialogue becomes self-contained atomic memories, retrieved through dense vectors plus a learned codebook’s co-occurrence structure. Online writing is separated from offline consolidation. This changes the retrieval unit and association mechanism rather than simply retaining more raw conversation. [3]

Decision: test associative retrieval against vector-only, lexical-hybrid and explicit-graph controls using the same evidence budget. Read Sections 3.2–3.5 and Appendix E.2: on a shared frozen memory store, the reported accuracy difference from HippoRAG2 was not statistically significant. Appendix A limits the work to conversational retrieval, not fully deployed tool-using agents; Appendix B excludes offline codebook training from per-dialogue runtime measurements. Treat it as a memory experiment, not a general GraphRAG replacement. [3]

## 5. Keep MCP runtime placement separate from its control plane

14 October 2026 · first-party engineering announcement · Apigee hybrid MCP support. Google describes tool configuration generated in its control plane and pulled into the customer’s Kubernetes cluster, where the MCP data plane handles requests. This extends its earlier managed-runtime offering to hybrid placement. It is optional and requires hybrid 1.17.0 or later. [4] [10]

Decision: teams already using Apigee can evaluate their existing API governance before building another gateway. Read the placement diagram and “Turning it on”; the enablement guide documents the Helm procedure and staged runtime rollout. The architecture still has a cloud control plane. This is vendor-documented availability, not evidence that a small business needs Apigee or that any deployment automatically meets its obligations. [4] [10]

Check two operational edges in the enablement guide: the data plane trusts the authenticated Message Processor, so other in-cluster access needs restriction. Failed configuration pulls retain the last known bundle indefinitely in hybrid 1.17.0; a Ready pod does not prove fresh policy. Monitor pull failures separately. [10]

## Turn one finding into a small, useful test

If you are handling sales, delivery and administration yourself, do not start by assembling a team of agents. Pick one recurring job whose acceptable result you can recognize. A reusable checklist or clearer input may solve it. Use the following worksheet only when a specific stage still needs help. These are proposed tests; none was performed for this edition.

Observed problem | Smallest useful comparison | Record before deciding
--- | --- | ---
Too much time discovering what to do | One capable agent versus the same lead with two planning workers; keep task mix and acceptance criteria fixed. | Accepted results, total tokens and charges, elapsed time, review minutes, and failed attempts.
Agents repeat each other’s reading | Current handoffs versus a bounded shared record on the same tasks. | Duplicate source reads, useful evidence recovered, incorrect reused claims, and failure recovery.
Memory misses related evidence | Vector-only versus a second retrieval path with equal evidence limits. | Supported-answer accuracy by question type, ingestion effort, retrieval latency, and maintenance.
The second tool turn fails | Replay the same sanitized transcript before and after a pinned runtime change. | Argument types, call IDs, output association, continuation, and invalid-input behavior.

Choose one row. Set a time and spending cap, save the baseline, and keep human approval for actions that reach customers. Include setup, checking, fixes and fallback work in the comparison. Stop if the result is less reliable or the extra coordination exceeds the benefit. Time released becomes a useful business result only when you decide what to do with it.

## Apple Silicon and local-runtime applicability

For an Apple Silicon workflow, begin with the exact runner, model artifact and API surface you actually use. The MLX-LM changes above do not establish behavior in Ollama, LM Studio, or exo; each needs its own replay. No Apple Silicon throughput, memory-capacity or distributed-inference test was performed here. A model running locally also does not make remote tools local.

## Window, exclusions and evidence limits

Reviewed 9–16 October 2026, following the 9 October edition. The three papers are fresh author-reported preprints; their experiments were not reproduced here. Release patches were read, and the MCP item is a first-party announcement. Older work and model additions without a stronger implementation consequence were not promoted as new findings. The dated prior edition remains available separately. Public search observations informed intent selection; traffic, customer demand and commercial uplift remain unknown.

## Sources

- [Scaling Inference-Time Compute with Teams of Coding Agents](https://arxiv.org/html/2610.17471v1) — [1] 12 October v1; Sections 2, 5–7; Appendix A and C.13; stage-dependent results and budget-comparison limits.; checked 2026-10-16
- [Scaling Multi-Agent Intelligence with Substrate-Centric Coordination](https://arxiv.org/html/2610.16120v1) — [2] 12 October v1; Sections 2, 3.4–3.5 and Appendix C; bounded reads, linked datastore, ablations and scope.; checked 2026-10-16
- [MemCode: Discrete Semantic Representations for Long-Term Agent Memory](https://arxiv.org/html/2610.19472v1) — [3] 14 October v1; methodology and Appendices A, B and E.2; same-store graph comparison and offline-training accounting.; checked 2026-10-16
- [Google Developers: MCP support in Apigee hybrid](https://developers.googleblog.com/announcing-mcp-support-in-apigee-hybrid-agentic-tools-that-run-in-your-own-cluster/) — [4] 14 October announcement; data/control-plane placement, hybrid prerequisites and staged rollout.; checked 2026-10-16
- [Ollama v0.40.5](https://github.com/ollama/ollama/releases/tag/v0.40.5) — [5] Release published 14 October at 19:45 UTC; custom-tool Responses continuation fix.; checked 2026-10-16
- [Ollama PR #18911: custom tool calls in Responses history](https://github.com/ollama/ollama/pull/18911/files) — [6] Inspected responses.go and responses_custom_tool_test.go; normalization, validation and upstream replay test. Not executed here.; checked 2026-10-16
- [MLX-LM v0.32.0](https://github.com/ml-explore/mlx-lm/releases/tag/v0.32.0) — [7] Release published 15 October at 18:30 UTC; argument parsing fixes. Subsequent patch releases exist; pin the exact tested version.; checked 2026-10-16
- [MLX-LM PR #1904: tool argument types](https://github.com/ml-explore/mlx-lm/pull/1904/files) — [8] Inspected server.py and test_server.py; only string arguments are JSON-decoded.; checked 2026-10-16
- [MLX-LM PR #1910: Qwen3 Coder untyped parameters](https://github.com/ml-explore/mlx-lm/pull/1910/files) — [9] Inspected parser and tests; parse valid objects/arrays without top-level type while preserving other strings.; checked 2026-10-16
- [Google Cloud: enable MCP in Apigee hybrid](https://docs.cloud.google.com/apigee/docs/api-platform/apigee-mcp/enable-mcp) — [10] Official enablement procedure, Message Processor trust boundary, and last-known configuration behavior on failed pulls; no deployment performed.; checked 2026-10-16