Thinking Machines Lab: Building Agentic AI and What It Means for Product Leaders

Nathan Binding

01 Jul 2025 — 1 min read

Overview

Mira Murati, former OpenAI CTO, launched Thinking Machines Lab in early 2025. In June, the startup closed a record $2 billion seed round, valuing it at $10 billion . Its mission: create agentic AI—systems that can both reason and take action—signalling a shift in how we define AI utility.

Moving from tools to actors

Traditional AI enhances features—autocomplete, search, recognition. Agentic AI, by contrast, acts on behalf of users: it plans, executes and adapts. That brings new requirements: context awareness, decision transparency and error handling.

Product implications

Defining agency: Start with clear boundaries. What decisions can the AI make? Where does human oversight sit?
UX design changes: Interfaces must reveal motivations and next steps. Users expect to inspect or intervene.
Safety infrastructure: Real-world actions require reliable fallback methods and rollbacks.
Ethical safeguards: Agentic systems can raise bias or misuse concerns. Audit trails and limits matter.

Roadmap adaptation

To support agentic AI, integrate long-range planning early. Combine feature-level goals with agent behaviour objectives. Build pipelines that feed models with feedback loops, and adapt metrics to include autonomy, trust and human collaboration.

Ecosystem and partnerships

Agentic AI needs access to external systems—APIs, user data, enterprise services. Partnerships become essential, covering domain expertise, compliance and integration tasks.

Challenges ahead

Complexity: Agents require architecture that manages planning, execution and monitoring.
Regulation: Expect frameworks on AI actions and accountability to evolve fast.
Skill shift: Teams need system designers, not just feature builders. Roles will change.

Thinking Machines Lab’s ambitions reflect a fundamental shift: AI will do more than assist—it will act. Product managers must rethink roadmaps, metrics, UX and team structure. Embrace agency now, and you stay ahead.

OpenAI ChatGPT Atlas: What I Expect After Using Arc, Dia and Comet

I have spent the last few months hopping between new browsers, trying to find one that fits how I actually work. I tried Arc, Dia, and Comet. Each one taught me something useful. None of them stuck. Arc looked promising at first. The design felt fresh and opinionated. Spaces and

We Made Software Easier to Use. Then We Took the UI Away

We’ve come full circle with interfaces We started with command lines. You sat in front of a black screen. You typed words you had to memorise. Miss one character and nothing worked. Power came from recall. Skill meant knowing the right command at the right time. Then desktops arrived.

My LoRaWan hobby

Introduction LoRa networks use low power long-range radio to send data across devices. When devices relay messages between each other, they form a mesh network. This can help you communicate across open spaces without depending on mobile signal, great for preppers or nerds. Meshtastic Meshtastic is an open source mesh

Digital IDs: What’s the Problem?

The new UK government digital ID scheme has sparked fierce debate. Critics warn of a surveillance state. But in truth, your personal data is already stored digitally at banks, government services, and private organisations every time you submit scans of passports, utility bills or bank statements. When you send a

Read more

OpenAI ChatGPT Atlas: What I Expect After Using Arc, Dia and Comet

We Made Software Easier to Use. Then We Took the UI Away

My LoRaWan hobby

Digital IDs: What’s the Problem?