Anthropic's Labs Lead On Fable's Capabilities + Building AI-Native Products — With Mike Krieger

Summary

Live from the Big Technology AI Summit, Mike Krieger explains how Anthropic Labs exists to keep products from lagging behind model capability gains and why that mandate has evolved as the company’s core offerings matured. He recounts early failures like unsafe computer use, then shows how more reliable delegation, dynamic workflows, and multimodal outputs made products like Claude Code practical. Krieger also argues that raw token consumption is a poor proxy for productivity, that outcome-based pricing is promising but hard for fuzzy knowledge work, and that Anthropic increasingly thinks about model launches in terms of intelligence, effort, and token efficiency together. On strategy, he says Anthropic must navigate being both a platform and a product while using transparency and shared building blocks to manage partners and competitors. He closes with a broader view that AI-native products will increasingly close the gap between what users intend and the actual work done, while safety and societal impact remain central.

Notes

• Anthropic Labs is an internal product org focused on frontier AI applications, not just model training.
• Krieger says he moved from CPO into the Labs role about five months earlier.
• He frames Labs as a way to keep products from falling behind exponential model improvements.
• Labs was originally created when Anthropic’s product engineering team was only about 25 people.
• The early models were useful but not strong enough for major delegation.
• He says day-of and week-of reactions to model releases are not reliable signals of true capability.
• For bio risk, Anthropic compares uplift from layperson use versus expert use or internet-only use.
• That framing treats safety as capability uplift, not just absolute model output.
• Krieger says the main product shift is broader delegation, not a simple level-up analogy.
• He emphasizes models remain spiky, inconsistent, and imperfect.
• The value of newer models is chunkier task delegation and better project-level planning.
• He describes a greater “theory of project,” where the model anticipates downstream dependencies.
• The labs mandate has shifted from pure product catch-up to aligning interfaces with future model behavior.
• Early computer use was unsafe because simple desktop-cleanup tasks could delete files.
• That failure mode blocked release until the product was iteratively improved and benchmarked.
• Claude Code artifacts were added because text-only responses were no longer enough.
• Anthropic is pushing multimodal output for richer explanations and visualizations.
• Krieger gives a large code-migration example: Python to TypeScript across hundreds of thousands of lines.
• Dynamic workflows can decompose work into subtasks, execution, verification, and double-checking.
• Earlier models often looked done while actually cutting corners.
• The newer models better preserve project context and avoid obvious shortcuts.
• Labs exists to explore what models will be good at next, not just what they are good at now.
• Claude Code emerged because no other product team had space to rethink coding form factors.
• The company now has thriving product lines, so Labs is less about basic catch-up and more about interaction design.
• Anthropic’s platform experience must evolve as model capabilities change.
• Krieger says Anthropic is both a platform and a product company.
• He says the dual role is something other companies, like Amazon, also have to navigate.
• Anthropic tries to manage partner tension through transparency.
• The company prefers shared building blocks such as plugins, skills, and MCP rather than closed ownership.
• Labs’ moonshot is giving Claude more agency and self-knowledge.
• The broader goal is to close the gap between user intent and the actual work performed.
• Giving Claude more environment context can make agent loops far more expressive.
• He says safeguards are essential when unlocking more environment access.
• Token count is a weak proxy for productivity.
• Anthropic evaluates launches on intelligence, effort, and token efficiency together.
• Outcome-based pricing is appealing, but fuzzy knowledge work makes it hard.
• Cloud Managed Agents can run infrastructure, agentic harnesses, and tool calls.
• The product supports both task mode and rubric-based outcome mode.
• Consumer AI breakout products have been limited so far.
• Platform consolidation and data gravity make consumer entry harder.
• He thinks Instagram could have been built with roughly four to six people using today’s tools.
• Modern AI would have helped more with platform parity than with pure headcount reduction.
• Anthropic rejects some products as likely bad for the world or morally fraught.
• Anthropic now uses economists to think about product impact on society.
• He is not working on compute or chips from Labs.
• He became more open to space-based data centers after a conversation about infinite power and land.