Armorer Guard: runtime control should start at the tool call

<!-- SC_OFF --><div class="md"><p>Been working with agents that can execute tools (shell, APIs, etc.) and noticed a gap:</p> <p>Unsafe behavior doesn’t show up in the prompt or

<!-- SC_OFF --><div class="md"><p>I've been thinking about this as local models get better at </p> <p>tool use — shell access, file system, HTTP requests, MCP. </p> <p>If a prompt

<p>Hosted agent runtimes are making a lot of infrastructure feel like a solved problem.</p> <p>Long-running sessions, context compaction, tool execution, sandboxes, and subagent

<p>Most API documentation is written for humans.</p> <p>MCP tool descriptions are different.</p> <p>They are read by the model that decides what to call next.</p> <p>That means

<blockquote> <p>Disclosure: I'm the author of <a href="https://github.com/razukc/capgate" rel="noopener noreferrer">capgate</a>, a compile-time policy compiler for MCP servers.

<p>I’m building AI CostGuard, a local-first TypeScript / Node.js runtime safety layer for AI agents.</p> <p>The problem I’m working on is not model quality or prompt

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s1g3dw/deterministic_agent_control_same_call_allow_then/"> <img alt="Deterministic agent control: same call