Tag
A discussion on the emerging security challenges when AI agents can take real actions (e.g., update records, call APIs) rather than just generate text, exploring needed infrastructure like scoped identities, policy layers, and action logging.
The author describes building a 'gate' that logs tool calls and verifies that claims in an AI agent's answer correspond to actual log entries, forcing the agent to show its work and reducing hallucinated responses.
A discussion on whether AI agent tool calls should be checked before execution, exploring safety and validation considerations.
New paper proposes self-compacting language model agents that can decide when to clean up their own traces of reasoning and tool calls to avoid accumulating mistakes and stale information.
A user reports that gpt-5.5 via the codex agentRuntime silently drops tool calls on short conversational turns, resulting in no response despite the model generating text. The issue is specific to gpt-5.5 and is fixed by switching to deepseek/deepseek-v4-pro, suggesting a regression.
Nanbeige 4.1, a 3B model from Chinese Indeed, outperforms larger Qwen models on tasks requiring 600+ tool calls.
Introduces Faramesh, an open-source runtime enforcement layer for AI agent tool calls that checks policies before actions run, offering a solution beyond observability or LLM-as-judge.
Kimi released K2.7 Code, a coding-focused AI model with improved benchmarks and 30% lower thinking token usage, emphasizing practical performance in long coding loops and agent tool integration rather than flashy scores.
An AI agent named Annie autonomously recompiled a Pokémon Ruby GBA ROM into a full hybrid WASM recompiler and GBA runtime, completing a task that would normally take an expert team months and cost tens of thousands of dollars.
The article argues that the primary failure point for AI agents in production is not the model itself, but the lack of infrastructure such as stop buttons, billing oversight, and traceability for tool calls.
The article presents a simple pattern to classify AI agent tools as safe or dangerous, routing dangerous actions like sending emails or deleting files to a human approval node to prevent unintended execution.
Setting temperature to 0 does not guarantee deterministic tool calls in agents due to batched inference causing floating-point reduction order shifts, leading to token flips and different actions under load.
Design Arena announces Agentic Game Development, a new multi-file, multi-turn evaluation with access to asset catalogs, built-in libraries like Howler and Tween.js, and expanded tool calls for sprite generation and asset discovery.
Cursor has released an auto-review mode that allows agents to execute tool calls with fewer approval prompts while maintaining safety.
The author shares lessons from instrumenting AI agent tool calls, revealing that tools like web_search can account for ~50% of spend, and highlighting the importance of tracking p95 latency and attributing costs per workflow or customer to avoid surprises.
LLC, a lightweight alternative to OpenWebUI for running local LLMs, updates to v0.6 with a chat converter to import OpenWebUI history and custom tool calls for user-defined tools.
A user running multiple agents reports that after upgrading to GPT-5.5, the model suddenly became less capable at executing tool calls and more prone to giving suggestions instead of acting, speculating OpenAI may be throttling for load management.