Blocking dangerous agent actions in prod — streaming is the guardrail blind spot
dank_as_fuck_ · reddit · 2026-09-27
The poster (founder of axonpush, disclosed) asks the community: once an agent holds real tool access (payments, CRMs, EHRs, databases), the risk is no longer a bad answer but a bad action — an oversized refund, a write to the wrong record, a tool it shouldn't touch. Most setups just log and review after the fact; he's polling whether people use hard checks inside tool functions, guardrails libraries, or human approval for risky actions.
He shares an honest engineering caveat: axonpush sits inline on OpenAI/Anthropic calls and can block a tool call by name or argument (e.g., amount > 1000) before execution, but this only works on non-streaming responses — with streaming, the tool call reaches your app before the full call is known, so it can only be flagged after the fact. He's also curious whether teams stream agent responses in prod or disable streaming for tool-heavy steps.
More from coding & agent
- Sydney Codex Community Buildathon set for Nov 14 with Cerebras-backed ultrafast compute track — paw_lean · 2026-11-11
- shadcn's improve hits 10k stars: let top models plan, cheap models execute code fixes — shadcn · 2026-11-11
- OpenAI open-sources Codex Security CLI and SDK for AI-driven vulnerability scanning and fixing — openai · 2026-11-11
- Claude Opus 5.5 runs an architecture search and finds AttnRes-Canon, a new top Transformer design — SonglinYang4 · 2026-11-11
- 5 research habits worth knowing: read papers backwards, reproduce first, wire coding agents via MCP — cneuralnetwork · 2026-11-11
- 'Adam can now do CAE': AI agent showcase in engineering simulation draws buzz — burhop · 2026-11-11