AI News Nuggets

Agent workflows need controlled retrieval and review that can stop a release

DeepSeek has opened separate Search and Fetch APIs for current web material, while Anthropic has described a layered review process for Claude Code changes. Both put a control point inside the workflow: decide what external content an agent may use, then require evidence and proportionate approval before its code ships.

Editorial read

This edition collects 2 notes across 2 topic areas and 2 sources. Start with Treat live web retrieval as a controlled tool call with source and access checks, Review agent-written code against runnable evidence and the exact commit being approved to get the week's main practical signal before scanning the remaining links.

Edition signal

The October 20 signal is to make outside inputs and release evidence visible

DeepSeek's October 17 documentation presents authenticated search results and single-page Markdown retrieval as separate JSON endpoints. Live retrieval is therefore a distinct dependency whose sources, access, costs, and failures need checking; a fetched page is still untrusted input. Anthropic's October 15 engineering account describes runnable checks before a pull request, automated bug and security review, human approvals tied to risk and a specific commit, and rollback after release. These are company-specific practices, not proven outcomes for every team. Keep source material traceable, let review effort follow the impact of the change, and name the person accountable for what ships.

ToolsSecurity

Tools October 20 note

Official API documentation

Treat live web retrieval as a controlled tool call with source and access checks

Source: DeepSeek

DeepSeek's October 17 update and current API guide document separate authenticated Search and Fetch endpoints. Search returns structured web results for a query; Fetch retrieves one web page as Markdown. The guide describes request limits, error handling, and cases where a target page cannot be fetched. DeepSeek's claims about search quality are vendor claims, and the announcement does not establish that retrieved content is accurate or safe for an agent to follow.

Why this matters: For an agent or retrieval-augmented workflow, decide when live web access is allowed and which sources matter. Record the query, returned URLs, retrieval time, and source used in an answer. Treat page text as data rather than instructions, bound spend and request rates, and handle blocked or failed fetches without inventing evidence. Test the workflow against representative tasks and primary sources before using its output for decisions.

Read DeepSeek's Search and Fetch guide

Security October 20 note

Official engineering article

Review agent-written code against runnable evidence and the exact commit being approved

Source: Claude Code team

Anthropic's October 15 engineering account describes executable checks and a fresh-context review before a pull request, automated bug and security checks on pull requests, risk-based human approval tied to a specific commit, merge-queue checks, staged release, and rollback. After three review rounds, its reviewer asks the author to reconsider the approach to avoid agent author and reviewer loops. This is Anthropic's own operating model, not a general performance guarantee.

Why this matters: Pilot the approach in one repository: require the agent to show test results and change intent, resolve substantive and security findings, and set human approval rules for sensitive changes. Recheck approval if the reviewed commit changes, cap repeated review passes, and measure reverts and incidents alongside reviewer time. Keep a named person responsible for release and recovery even when most checks are automated.

Read the Claude Code team's review process