The Sequence
8/10 signal
The Sequence Radar #893: Last Week in AI: GPT-5.6, Grok 4.5, Muse Spark 1.1 and the Post-Chatbot Stack
agentictool-usemodelsmulti-agent
What happened
Recent model announcements, including the hypothetical GPT-5.6, Grok 4.5, and Muse Spark 1.1, signal a fundamental shift from chat interfaces to agentic execution runtimes. Key architectural innovations include programmatic tool calling, where models write code to coordinate tools, parallel subagent execution, and full-duplex audio architectures for real-time interaction.
Context
The dominant paradigm for interacting with LLMs has been the chat interface, a simple request-response loop. While effective for Q&A and text generation, this model is limited for complex, multi-step tasks. The introduction of function calling was a step towards agency, but it still relied on a rigid, often stateless, JSON-based contract. The industry is now moving to address these limitations by building models and platforms that function as persistent, stateful runtimes. This new "post-chatbot stack" aims to support long-running, multi-application workflows and more natural, real-time human-computer interaction, moving beyond the simple text box.
Key points
- GPT-5.6 reportedly introduces programmatic tool calling, a significant evolution from standard JSON-based function calling. Instead of just selecting a tool and its parameters, the model writes and executes small programs or logic loops to coordinate multiple tools locally, enabling more complex and stateful task execution before returning a final result.
- The architecture of these new systems incorporates parallel subagents. This allows a primary agent to delegate tasks to specialized subagents that can run concurrently, improving efficiency and enabling more complex problem decomposition. This moves from a monolithic model to a distributed, multi-agent system.
- GPT-Live introduces a full-duplex audio architecture. This allows the system to listen and speak simultaneously, enabling more natural, interruptible conversations. It delegates reasoning tasks to a background process while maintaining the real-time audio stream, fundamentally changing the design of conversational state machines.
- ChatGPT Work represents a shift towards long-running, cross-application agentic workflows. The focus is on agents that can operate persistently across a user's connected files, websites, and applications, moving beyond single-session, sandboxed interactions to become an integrated part of a user's digital environment.
What's new
The core novelty lies in treating the LLM as an orchestration engine or runtime, not just a text generator. Programmatic tool calling is a major architectural shift from declarative JSON function calls, allowing for imperative logic execution. Similarly, a full-duplex architecture for real-time audio is a fundamental change from the traditional turn-based, request-response model of conversational AI.
Limitations
The analysis is based on unconfirmed or hypothetical product names and features like 'GPT-5.6' and 'Grok 4.5'. The described capabilities are forward-looking and may not reflect the actual state of any currently released products. The information is speculative and derived from high-level analysis rather than direct technical documentation.
The take
The metaphor for LLMs is rapidly shifting from 'autocomplete in a text box' to a 'distributed operating system'. Programmatic tool calling is a massive upgrade, allowing models to run local logic loops and manage state without constant round-trips to a central controller. This is far more powerful than simple function calling. Likewise, full-duplex audio isn't just a feature; it's a paradigm shift that forces a complete rethink of how to design and manage state in real-time conversational agents. We are witnessing the foundational components of truly agentic computing being built directly into frontier models.
Signal
The future of AI interaction is moving beyond the chat window. The focus is shifting to building persistent, agentic runtimes that can execute complex, long-running tasks across multiple applications. This signals a move toward AI as an integrated, ambient layer of computing rather than a standalone destination application.
Don't read this site daily. Get it in your inbox.
The daily brief and Sunday deep dive — distilled, scored, and opinionated. For builders only.