Tool Use & Environment

23 篇文章

Agent-First Tooling and Logging

Most developer tools, CLIs, and application logs are designed for human consumption. They use color-coded, multi-line, or summarized outputs that are easy for a person to scan but can be difficult for

emerging

Agent SDK for Programmatic Control

Interactive terminal or chat interfaces are suitable for many agent tasks, but not for all. Integrating agent capabilities into automated workflows (e.g., CI/CD pipelines, scheduled jobs, batch proces

emerging

Agentic Search Over Vector Embeddings

Vector embeddings for code search require: - Continuous re-indexing as code changes - Handling local uncommitted changes - Additional security surface area for enterprise deployments - Infrastructure

best-practice

AI Web Search Agent Loop

Traditional LLMs have a training cutoff date, meaning they don't know recent facts or real-time information. Simply connecting a model to a search API isn't enough - the model needs to: - Decide when

emerging

CLI-First Skill Design

When building agent skills (reusable capabilities), there's tension between: - **API-first design**: Skills as functions/classes—great for programmatic use, but hard to debug and test manually - **GU

emerging

CLI-Native Agent Orchestration

Web chat UIs are awkward for repeat runs, local file edits, or scripting inside CI pipelines.

proposed

Code Mode MCP Tool Interface Improvement Pattern

Traditional Model Context Protocol (MCP) approaches of directly exposing tools to Large Language Models create significant token waste and complexity issues. We've moved from telling LLMs what to do,

established

Code-Over-API Pattern

When agents make direct API or tool calls, all intermediate data must flow through the model's context window. For data-heavy workflows (processing spreadsheets, filtering logs, transforming datasets)

established

Code-Then-Execute Pattern

Plan lists are opaque; we want **full data-flow analysis** and taint tracking.

emerging

Dual-Use Tool Design

Building separate tools for humans and AI agents creates: - **Maintenance overhead**: Two implementations of similar functionality - **Inconsistent behavior**: Human tools work differently than agent

best-practice

Dynamic Code Injection (On-Demand File Fetch)

During an interactive coding session, a user or agent may need to inspect or modify files **not originally loaded** into the main context. Manually copying/pasting entire files into the prompt is: -

established

Egress Lockdown (No-Exfiltration Channel)

Even with private-data access and untrusted inputs, attacks fail if the agent has **no way to transmit stolen data**. Many real-world fixes simply removed or filtered outbound channels.

established

Intelligent Bash Tool Execution

Secure, reliable command execution from agents is complex and error-prone: - **PTY requirements**: TTY-required CLIs (coding agents, terminal UIs) fail with direct exec - **Platform differences**: Li

validated-in-production

LLM-Friendly API Design

For AI agents to reliably and effectively use tools, especially APIs or internal libraries, the design of these interfaces matters. APIs designed solely for human consumption might be ambiguous or ove

emerging

Multi-Platform Communication Aggregation

Users communicate across multiple platforms (email, Slack, iMessage, etc.) and need to search for information that might exist in any of them. Searching each platform manually is slow and error-prone.

emerging

Multi-Platform Webhook Triggers

An internal agent only provides value when its workflows are initiated. Building out a library of workflow initialization mechanisms (triggers) is core to agent adoption. Without easy triggers, employ

emerging

Patch Steering via Prompted Tool Selection

Coding agents with access to multiple patching or refactoring tools (e.g., `apply_patch`, `AST-refactorer`, `codemod`) may choose suboptimal tools if not explicitly guided. This leads to: - **Unneces

best-practice

Progressive Tool Discovery

When agents have access to large tool catalogs (dozens or hundreds of available tools), loading all tool definitions upfront consumes excessive context window space. Most tools won't be used in a give

established

Shell Command Contextualization

When an AI agent interacts with a local development environment, it often needs to execute shell commands (e.g., run linters, check git status, list files) and then use the output of these commands as

established

Subagent Compilation Checker

Large coding tasks often involve multiple independent components (e.g., microservices, libraries). Having the **main agent** handle compilation and error checking for every component in-context: - **

emerging

Tool Use Steering via Prompting

AI agents equipped with multiple tools (e.g., shell access, file system operations, web search, custom CLIs) need clear guidance on when, why, and how to use these tools effectively. Simply having too

best-practice

Virtual Machine Operator Agent

AI agents need to perform complex tasks beyond simple code generation or text manipulation. They require the ability to interact with a full computer environment to execute code, manage system resourc

established

Visual AI Multimodal Integration

Many real-world tasks require understanding and processing visual information alongside text. Traditional text-only agents miss critical information present in images, videos, diagrams, and visual int

emerging