Your agents finished the job. What did the bad answer touch?

011. Containment now follows the data, not only the agent

Airlock separates authority from influence. A delegation tree narrows permissions from parent to child, while an append-only provenance graph records which executions consumed each result. When an alert arrives, the system can stop affected downstream actions without freezing unrelated work. [1]

The preprint reports 1,832 testbed runs, 161 physical smart-home LLM runs and 24,266 simulation records. Its main grids report no post-alert faulty actuations after containment converged. Those are author-reported results on the published runtime, not evidence that the same guarantees carry into another framework or deployment.

Decision affected: before adding another autonomous worker, ask whether you can trace what it read and revoke the exact external capability its output reached. Read Sections 3 and 4 for the authority tree, provenance DAG and fence-then-prune design; read Section 5 and Appendix B for the comparison grids and runtime artifact.

022. An MCP server can schedule scarce machine capacity

ToolSched moves shared-host coordination into a shell MCP server. Each command arrives with an estimated memory, CPU and I/O demand. A queue, selector, sampler and controller combine those estimates with live host conditions, stagger execution and restore from checkpoints after resource errors. [2]

On Terminal Bench, the authors report 148.4 percent higher resource utilization and 29.4 percent higher command throughput than unscheduled execution. The unscheduled baseline also failed 11 tasks, so the similar total runtime is not a like-for-like completion result. The experiment targets Linux containers on a shared host, not Apple Silicon process scheduling.

Decision affected: if several agents compete for one workstation, measure queue delay, completed tasks, peak memory and recovery cost before buying capacity. Read Section 3 for the queue and correction loop, Table 2 for completion and runtime, and Section 4.4 for checkpoint overhead.

033. Graph retrieval can discover the interface at query time

ArcRAG tests whether retrieval can survive a change in graph construction. At inference time it inspects the labels, relations and properties reachable from named entities, plans a multi-hop traversal, generates Cypher and repairs failed queries. The same configuration is evaluated against schema-guided and open-information-extraction graphs. [3]

This changes the migration question. A team may not need to bind retrieval logic to one graph-building pipeline, but construction still changes absolute accuracy. The paper compares frozen graph families across MuSiQue, HotpotQA and TriviaQA; it does not establish portability to a private ontology, access-control layer or production latency budget.

Decision affected: test one unchanged retriever against two representations of the same corpus before rebuilding the whole knowledge layer. Read Sections 3 and 4 for construction and interface induction, then the main results and ablations for retry and pass budgets.

044. Weight memory is useful only with a route back to evidence

Route to Remember trains a per-user LoRA with on-policy self-distillation, then audits its context-free answer against a text store. If the answer conflicts with retrieved evidence, the system regenerates with the evidence in context. The point is not that weights replace retrieval. The two stores fail differently. [4]

The paper reports 59.9 percent accuracy for its weight memory and 66.1 percent after verify-then-regenerate on its main protocol. Older content decays in the adapter, while retrieval remains comparatively stable. The implementation trains 30 rounds per session and uses a separate judge, so write cost, privacy and reproducibility need a local evaluation before adoption.

Decision affected: reserve parametric memory for frequently reused facts or habits only if a text store can still verify absence, conflict and updates. Read Sections 3.3, 4.2 to 4.4, Appendix A.5 and Tables 7 to 10.

055. Local runtime upgrades can change the conversation contract

Ollama 0.40.14 stops OpenAI-compatible endpoints from replacing packaged sampler defaults when clients omit temperature or top-p. It also preserves readable input in the Responses API, retries ChatGPT disconnects, validates Codex routing and accelerates top-k sampling on MLX for large vocabularies. [5]

This is production release evidence, not a benchmark. For Apple Silicon users, faster MLX sampling is relevant, but the more important testing change is behavioral parity: an upgrade can alter defaults, replayed input and recovery even when the model file is unchanged. LM Studio and exo applicability remains unknown from this Ollama release.

Decision affected: pin the runtime, preserve request parameters and replay one multi-turn tool conversation before and after an upgrade. Read the v0.40.14 release notes and the linked changelog, then inspect the exact client path you use.

06A containment worksheet before you add another agent

Swipe or scroll to compare

For a business owner, the immediate choice is smaller: do not multiply automated work until one workflow can show what it was allowed to do, what it touched and how a person can stop or correct it. A clear process, template or single assistant may remain cheaper to review and maintain.

07Window and evidence limits

This brief covers primary material published from 3 to 4 November 2026 and the Ollama release dated 3 November, checked on 6 November. Paper claims were inspected but not reproduced. No private infrastructure, customer result, local benchmark, production use or ranking uplift is claimed.