Technical Articles

Deep technical notes focused on the contracts that make AI systems reliable in production: realtime interaction, asynchronous work, tool execution, voice timing, observability and architecture trade-offs.

Realtime agents and tool timing

• Proactive and reactive agents and tool calls — why a model turn should not be the clock that governs asynchronous tools; acceptance, durable work and proactive delivery as separate state machines.
• Reactive–proactive voice agents — how to keep a spoken conversation alive while slow work continues, then deliver results without interrupting the user or duplicating effects.
• Three architectures for voice agents — full cascade vs half cascade vs speech-to-speech, including prosody, latency, barge-in, tools and the hybrid S2S-surface architecture.

Reading order

Start with Proactive and reactive agents and tool calls for the general asynchronous runtime. Continue with Reactive–proactive voice agents for the interaction/delivery problem. Finish with Three architectures for voice agents to compare the modality stack itself.

The articles intentionally separate technical completion from human-visible delivery. That separation is the recurring design rule across all three notes.