Matt Okner

Runtime guardrail for OpenAI Agents SDK tools. It checks tool names, schemas, and arguments, then allows, blocks, or asks for approval before side effects happen.

• Tool boundary: Wraps an agent's tool array without changing the agent loop
• Policy engine: Allows reads, gates writes and external sends, blocks secrets and payment-like calls
• Approvals: Pauses risky actions for Telegram or custom review
• Audit trail: Writes JSONL decisions with redacted arguments
• Setup: Scaffold a guarded agent or wrap existing tools with guardTools()

Compared GPT-4o character counting with direct reasoning versus a Python interpreter. Tool use removed the measured counting errors.

• Baseline: 77.8% accuracy without tools on repeated or long strings
• Tool run: 100% accuracy with a Python interpreter
• Finding: Token boundaries cause avoidable counting mistakes
• Harness: Reproducible eval prompts and result checks

Voice AI for inbound and outbound calls where response time and reliable handoff matter.

• Latency: Targets sub-500ms speech-to-speech turns
• Transport: Twilio Media Streams over WebSockets
• Speech: Deepgram STT with low-latency synthesis
• Agent loop: Routes intents and tool calls through Anthropic
• Deploy: Runs on Railway with graceful session cleanup