Agent tools are being built around proof, scoped memory, and accountable action

GlueRun-go makes parallel coding agents easier to inspect. Each task runs in its own Git worktree, holds a JSON lease, and writes a state packet covering owned files, changed files, commands, tests, and evidence. The auditor checks that packet and the gate result before the system retries, narrows scope, escalates, or parks the task. The reported gains are operational: detached dispatch returns reconcile within seconds, and crash detection drops from a 60-minute stale-lease window to about one reconcile cycle.

Codeflowmap attacks the review side of the same problem. It builds import and TypeScript/JavaScript call graphs through static analysis, then lets an optional large language model (LLM) add per-file notes for reads, writes, config, auth, and flows. That keeps graph edges tied to deterministic analysis while using model output for semantic annotation. The project reports no benchmark or accuracy score, so its value is in the inspection workflow it exposes.

Two projects focus on memory that agents can query during work. Vitrus keeps company knowledge in Markdown, builds a disposable index, and returns sourced answers with confidence, freshness, and deterministic gaps. Its think and verify commands classify answers as grounded, stale, contradicted, or unsupported. The repo reports source-hit of at least 90%, perfect gap recall and precision on a controlled synthetic corpus, zero unauthorized results in leak testing, and more than 200 tests.

Callimachus targets coding-agent history. It imports threads from 11 sources into a local SQLite store, mixes keyword and vector retrieval, and exposes the index through desktop, command line, VS Code/Cursor, and Model Context Protocol (MCP) tools. Its concrete evidence is coverage and speed: 16 MCP tools, 21 command-line commands, local 384-dimensional embeddings, and about 90,000 Claude messages indexed in roughly 25 seconds. It does not report retrieval accuracy or a user study.

FERNme proposes user-owned preference memory for action-taking agents. It stores sparse per-site graphs, updates edge weights with deterministic co-occurrence rules, and retrieves a short memory card for the agent prompt. The design reduces per-turn LLM calls and gives users inspection, edit, export, and delete controls.

The strongest reported numbers come with clear limits. The project reports 88 tests, a +0.06 precision@5 cold-start gain during turns 1–3, a +16% relative conversion lift in a simulated storefront, and modeled quality of about 80–90% of an LLM ceiling at one to two orders of magnitude lower cost. The author states that the main evidence uses synthetic or LLM-authored data, and the real-human pilot is still pending.

Amazon’s position on agent governance centers on accountability and permission scope. The article argues that repeated human approvals degrade under high-volume, low-signal review. Amazon’s preferred control model gives each agent its own identity, logs the agent and the human it acted for, and limits permissions by task risk.

The same control pattern appears in Vitrus at the tool boundary. Its OpenAPI import, search, verify, and call flow checks endpoint names, arguments, types, deprecated endpoints, and permissions before execution. Both cases treat approval as one control among several, with audit identity and pre-execution checks carrying much of the burden. The Amazon article gives no benchmark or safety metric, so the claim is a governance design argument backed by concrete failure examples.