- NATURAL 20
- Posts
- Anthropic, Microsoft, and xAI Push AI Agents Beyond Simple Chatbots
Anthropic, Microsoft, and xAI Push AI Agents Beyond Simple Chatbots
PLUS: Terence Tao outlines Math 2.0, OpenAI updates Codex and dots, and Anthropic expands AI-powered cyber defense.
Wes Roth
October 12, 2026
In partnership with
Agents now issue refunds, update accounts, and close tickets end to end. Running Agents in Customer Work: four Tuesdays on agentic AI in customer ops. Register for All Four
Today:
Claude Managed Agents opens dynamic multiagent workflows in public beta
Grok Bot gains a dedicated inbox for communicating with outside services
Microsoft introduces Decision-1 for fast structured scoring
Terence Tao outlines a vision for AI-assisted Math 2.0
Codex suggests follow-up instructions with Composer predictions
From Answering Questions to Coordinating Work
AI releases are increasingly focused on what happens between a user's request and the finished result. Anthropic is letting a lead agent plan and run groups of other agents. Grok Bot now has an inbox for communicating with services and people. Microsoft has introduced a specialized model for the fast decisions that help control longer workflows.
Taken together, these developments show why agent infrastructure is becoming a separate layer of the AI stack. Coordination, external communication and low-latency classification each solve a different problem that a single chatbot response cannot address.
Anthropic opened dynamic workflows in Claude Managed Agents to public beta, adding a way to coordinate many agents without manually defining every handoff. A lead agent writes an orchestration program, distributes pieces of a complex assignment across other agents and combines their results. The feature is designed for work such as codebase reviews, migrations, large research tasks and audits where separate steps can run in parallel.
The system offers several forms of assistance. A lead agent can delegate to subagents that retain their own conversation threads, run a programmed workflow in the background or consult an advisor model. In a workflow run, agents work in isolated threads, and the server handles the orchestration while the primary session can continue communicating with the user.
Developers configure multiagent behavior in an agent definition and can allow predefined specialists, agents defined during execution or both. Anthropic advises using session budgets and clear instructions about when to invoke workflows because running many agents can consume substantial tokens.
The new orchestration options provide more structure for long jobs but do not guarantee that every task benefits from parallelism. Workflows are best suited to assignments with independent steps, while tasks that require a tightly coupled sequence or frequent direction from one agent may need a simpler approach.
Grok Bot can now claim a dedicated email address at mail.grokbot.com, giving agents an inbox separate from the user's personal or work email. The feature can help a Bot register for services, receive verification messages with approval, contact a business, manage replies and collect records such as newsletters or receipts.
According to the product's help documentation, an individual account can claim one address, and all its Bots may use that address in the owner's chats or approved routines. A user can ask a Bot to create an address or claim one through the Email plugin. Team administrators must enable the capability before members can use it.
Incoming mail can trigger a routine only when the user has configured that routine. Sending mail also requires a user request or an approved routine, and instructions inside a received message are not treated as authorization to act. The feature does not provide a conventional inbox interface; users ask their Bot to search or summarize incoming messages.
An email address extends what a persistent agent can do outside its own app, but it also increases the importance of access controls, sender verification and review of outbound messages. The feature is rolling out, with availability and team controls described in the official help documentation.
Microsoft released Microsoft-Decision-1, a model designed to make fast, structured decisions rather than produce open-ended text. It returns probability scores for predefined choices, making it suitable for classification, routing, prioritization, verification and workflow control. The model is available through Microsoft Foundry, with OpenRouter availability planned.
Microsoft says Decision-1 is based initially on a trained Qwen3.5-9B model and supports yes-or-no questions, multiple-choice choices, ratings and rubric-based evaluations. Its output can help an application choose the next action or evaluate an agent's work without asking a general-purpose language model to generate a long response.
In Microsoft's evaluation across 36 benchmarks and nearly 150,000 questions, the model achieved the highest overall accuracy among the systems it compared. The company reports median latency about 4.5 times faster than the next-fastest model, Quyet-1.0-Large, and about 35 times faster than GPT-6 Sol. Those are vendor-reported comparisons under Microsoft's test conditions, not independently established gains for every application.
Microsoft lists pricing of $0.042 per million input tokens with no output-token charge. The release adds another option to the emerging category of decision-focused models, where predictable choices and low latency can matter more than conversational fluency.
🧠RESEARCH
Researchers presented TokenRouter, a serving system that routes work between language models at the token level rather than assigning an entire question to one model. It uses asynchronous per-model subservers and delayed batching to address desynchronized generation steps and admission delays. Across the studied model pairs, workloads and routing strategies, the authors report decoding throughput 2.01 to 64.15 times higher than the comparison systems; the result is specific to their evaluation and should not be read as a universal production speedup.
AgentGarten combines game engines or simulators, which maintain consistent world rules, with a shared neural renderer that produces realistic visual observations. Agents explore these environments, review what happened and pass written playbooks to later rounds, allowing behaviors to improve without retraining the underlying model each time. The authors describe strategies emerging after several rounds, but their comparison with conventional reinforcement learning uses different starting conditions and does not establish a direct, like-for-like learning-speed advantage.
Learn2Play Bench evaluates language-model agents on new text games whose rules must be discovered through interaction rather than recalled from training or supplied instructions. The researchers vary models, agent frameworks and experience retention while testing whether learned behaviors transfer to new instances. They find that retaining full interaction histories can work better than compressing experience into summaries, that top human players achieve higher peak scores than tested agents, and that the chosen agent framework can improve results while lowering estimated inference cost.
📲SOCIAL MEDIA
🗞️MORE NEWS
Mathematician Terence Tao published slides from his October 9 Caltech lecture, Math 2.0, describing how the growth of automated mathematical problem-solving could change research as proofs become easier to generate. He argues that mathematical work must still preserve human understanding, exposition, community collaboration and a connection to real-world applications instead of treating the production of proofs as the only objective. His proposal looks toward AI tools that help sustain the mathematical community and improve how mathematical results are interpreted and used.
OpenAI introduced Composer predictions in beta for the Codex desktop app, where the software can suggest a user's next message after Codex finishes responding. The feature works in local and SSH threads with GPT-6 Astra or GPT-6.1 Sol for eligible adult personal Pro subscribers, and pressing Tab accepts a suggestion for review or editing rather than sending it automatically. During beta, generating suggestions does not consume Codex credits or limits, although sent messages count toward normal usage.
OpenAI updated dots so eligible users can create and customize their dot directly in the ChatGPT mobile app on iOS and Android, including connecting plugins during setup. A dot can now initiate Codex work, follow up on existing coding threads and use context from related ChatGPT conversations, Codex work and automations. Dots can also read and edit scheduled ChatGPT Work tasks, with access continuing to roll out gradually according to plan, region and workspace settings.
Anthropic launched its Cyber Mission to support the defense of critical infrastructure and open-source software, including power grids, water systems and transportation networks. Its Critical Infrastructure Defense Program will provide frontier models, on-site engineers and threat research to trusted organizations, while an opt-in OSS Scanner offers open-source projects recurring vulnerability scans at no cost. The scanner's findings are generated without human review, so maintainers may still need to confirm reports before treating them as valid vulnerabilities.
What'd you think of today's edition? |
Reply