Introducing Tool Retrieval in the Gemini API
Preview in the Interactions API: add a tool_search tool with defer_loading true and Gemini fetches parameter schemas only when needed, cutting input tokens for large tool libraries.
23 articles · 6 categories
The finishable daily brief
Thursday, Nov 19, 2026
23 articles · 6 categories
read top to bottom · then stop
In 30 seconds
The tool layer moved out of the prompt today. Gemini now loads tool schemas on demand, LangSmith put credentials and permissions behind a gateway, and Vercel shipped eve 1.0.
Security evidence cut both ways. An audit found AgentDojo scored 433 real attacks as misses, while an autonomous Codex run reported 507 Webmin bugs for about $4,578.
Prefer to watch?
Vendors moved tool handling out of the prompt: schemas load on demand, credentials sit behind a gateway, and output JSON gets compressed in a proxy.
Preview in the Interactions API: add a tool_search tool with defer_loading true and Gemini fetches parameter schemas only when needed, cutting input tokens for large tool libraries.
Public beta managed tool layer that holds credentials, decides which tools each agent may call and under whose identity, and logs every action.
Zero-dependency Go reverse proxy that rewrites repeated JSON keys into positional tuples, claiming 35–60% fewer output tokens for structured outputs and function calls.
Solo project that refuses tool calls when untrusted prompt text tries to mint authority or replay receipts; publishes 3,315/3,315 verified receipt hashes.
Agent runtimes are becoming infrastructure with GA labels, in-house standards, and durable state, not per-team experiments.
Vercel's open-source agent framework hits GA after 230+ preview releases, adding approval controls for actions and an eval suite for how well coding agents build with it.
HubSpot explains why it built one internal agent harness in 2025 after teams kept making divergent model and architecture choices.
Hatchet argues agent sandboxes need coordinated snapshots of state across processes, not just isolated compute.
Docker docs now lead with agent sandboxes: scheduled tasks run in a fresh cloud sandbox that is deleted when the task ends or times out.
Durable, event-driven MCP message bus with bundled SQLite so several coding agents can hand off API changes and test failures without a human relaying them.
Today's eval work audits actions and traces instead of trusting a model's reply text or one automated verdict.
Audit of AgentDojo: the benchmark scored 433 runs as no attack even though the agent paid the attacker or took another decisive action, 11.9% of real attacks. Preprint, not peer reviewed.
Open-source, offline suite of 165 deterministic tests over OpenTelemetry GenAI traces that flags unearned success claims, hidden-instruction follows, and forbidden tool calls with no LLM judge.
Audiala swapped an LLM router for a decision model and rolled back in 23 minutes: the old call also resolved cities, references, and tool flags nobody evaluated.
Databricks used 125 insurance questions to show agents need governed business definitions; Finance and Underwriting gave opposite, correct loss-ratio answers.
Autonomous auditing got cheap enough to run on a whole codebase while package and model supply chains stayed the soft spot.
A fully autonomous Codex campaign reported 507 reproduced vulnerabilities in Webmin 2.670 using 8.35B tokens, about US$4,578 at standard API rates.
Follow-up: PyPI quarantined the Microsoft publisher because the attacker is believed to still have publishing access; treat any 1.26.1 or 1.26.2 install as a credential compromise.
Dragos argues for canary tokens across the AI supply chain so a leaked key or poisoned dependency trips an alarm when used.
Large operators described agents that run inside their own engineering loops, from performance fixes to RL rollouts.
OpenAI runs Codex-powered workflows that find and fix performance regressions to keep ChatGPT fast.
Meta uses agents to port proven ranking techniques across its many ads models, each with different architecture, data, and serving limits.
Abundant runs RL rollouts for frontier labs on Modal Sandboxes, giving each task its own isolated environment, Dockerfile, and verifier.
Chinese and Indian labs competed on price while compute commitments shifted in both directions.
DeepSeek no longer charges separately when the model triggers search or web fetch, which simplifies cost math for tool-using agents.
Sarvam's new token factory lists its own model below DeepSeek's price.
Moonshot AI cancelled a $150 million Nvidia B300 GPU rental deal in Indonesia.
At Ignite, Microsoft widens access to Azure VMs on NVIDIA Vera Rubin NVL72 systems and extends the stack to Windows PCs.
You are caught up for this edition