When Your Model Outpaces Your Workflow: Mike Krieger on Building with Claude Fable 5

Mike Krieger, head of Anthropic Labs and Instagram co-founder, says Fable 5 did not just speed up his existing workflow. It made those workflows irrelevant. The main point: the bottleneck has shifted from execution to verification. The competitive advantage now goes to people who can trust models to work autonomously over hours, not those who can prompt faster. This analysis is for technical leaders who want to understand where the real leverage is, and where common wisdom about AI coding tools is already outdated.

The New Rhythm: From Turn-by-Turn to Overnight Delegation

The biggest change Krieger describes is not about code quality or speed. It is about the fundamental rhythm of work. When he first got access to Fable 5, his reaction was telling: "I feel like a total newbie again because I feel like the way that I am prompting or even thinking about decomposing a task is really out of date now with this model."

This is not modesty. It is an observation about what happens when a tool's capability crosses a threshold. The old model was conversational: you prompt, it responds, you refine. That workflow assumed the model needed constant guidance. Fable 5 changes that assumption.

Krieger now routinely sets Claude working on complex tasks overnight. The model does not just complete the work. It handles unexpected failures autonomously. "I was like, okay, all right, well, my task is to do this complex task overnight. I got stuck because this remote service went down. I'm gonna write a scaffolded backend for now so I'll document that, you know, go all the way through."

This creates a new kind of leverage. The model does not just execute. It adapts. It makes judgment calls about when to fix a problem versus when to work around it. Krieger describes waking up to find Claude had completed the work by 2 AM and then "just photos its thumbs for the next four hours." The implication is clear: the model has developed a sense of completion, not just correctness.

"I think the most impressive thing for me is you're just being able to delegate that kind of level of task and just trust that the right thing will happen by the end."

-- Mike Krieger

The Cost Collapse Nobody Is Modeling Correctly

Krieger draws a direct comparison between building with Fable 5 and building Instagram's first version. "Instagram v1... was like five days of all-nighters be working on the front end and back end and Kevin working on the initial filters to get that out." The media tracker he built over a weekend with Fable 5 was not just faster. It was architecturally different. It could modify itself from within, a capability that would have been unthinkable in 2010.

But here is where the systems thinking gets interesting. The cost to build has collapsed, but the cost to verify has not followed the same curve. Krieger notes that he has become "verification killed" -- spending significant energy on ensuring the model's output is correct. This creates a hidden dynamic: the faster you can build, the more you need robust verification loops, or you are just generating technical debt at machine speed.

The pricing question adds another layer. Fable 5 is expensive, and Krieger admits he has become more thoughtful about usage. But he reframes the cost calculation: "I think measuring cost has gotten so multifaceted now because there is the per turn costs. And then there was like, what did it cost you not to just do the task, but like complete the task to your satisfaction?" The model's ability to get it right in fewer turns means the total cost of ownership may actually be lower, but only if you are measuring the right thing.

The Death of the Solo Coder and the Rise of the Orchestrator

Krieger's answer to "is software engineering over?" is nuanced but revealing. "Software engineering is different. It has dramatically changed." The craft of typing code, dreaming about elegant solutions, and debugging at 3 AM is passing. But the higher-order skills are becoming more valuable.

What does this look like inside Anthropic? Engineers now manage multiple Claude sessions simultaneously. They have built personal dashboards to track what their AI agents are doing. The role has shifted from writing code to orchestrating work, maintaining context, and making judgment calls about architecture and trade-offs.

"I think there is a feeling of loss I think in some of the better engineers that I talk to as well as the feeling of oh my God but I can do insane amounts of work now at the same time. So we're holding both ideas in our heads at once."

-- Mike Krieger

This is the systems-level insight that most commentary misses. The transition is not pain-free. Engineers who loved the craft of coding are grieving. But the ones who adapt are finding they can do work that was previously impossible. Krieger's example of porting a Python codebase to TypeScript over a weekend using dynamic workflows, something he says would have been unthinkable at Instagram, illustrates the new ceiling.

The Verification Frontier

The most practical insight from this conversation is about verification. Krieger's approach has evolved into a multi-layered system: screenshots and video captures for visual verification, mock backends for integration testing, and adversarial testing within dynamic workflows. The key innovation is treating verification as a first-class concern, not an afterthought.

He describes giving Claude video captures of its own output and watching it scrub through to identify animation jank. "I never would be able to do it with a screenshot latency capture because we will have missed the moment." This is a new kind of feedback loop -- the model verifying its own work using modalities humans cannot easily process.

The deeper point is about trust. Krieger now sets Claude working on tasks with the confidence that it will handle edge cases. But that trust is earned through systematic verification, not blind faith. The teams that will win are the ones that build verification infrastructure as carefully as they build product features.

Key Action Items

• Build verification infrastructure early. Screenshot galleries, video capture, and mock backends are not optional. They are the prerequisite for trusting autonomous model work. Invest in these systems over the next quarter before scaling your model usage.
• Redesign your workflow around overnight delegation. The biggest leverage comes from tasks that take hours, not minutes. Start by identifying one complex task per week that you can set running before you leave. This pays off in four to six weeks as you build confidence.
• Create personal dashboards for agent orchestration. If you are managing multiple Claude sessions, you need a way to track what is happening. Build a simple dashboard that shows active sessions, pending reviews, and completed work. This is an immediate action that compounds over time.
• Separate planning from execution. Use Fable 5 for architectural conversations and planning, then delegate execution. Krieger's approach of having the model produce diagrams and markdown documents for team alignment before building is a pattern worth adopting immediately.
• Accept the emotional transition. The grief over losing the craft of coding is real and normal. Acknowledge it in your team. But do not let nostalgia prevent you from capturing the upside. The discomfort now creates advantage in 12 to 18 months when your competitors are still trying to figure out how to prompt effectively.
• Measure total cost of ownership, not per-turn cost. Fable 5 is expensive per token but may be cheaper per completed task. Build a simple tracking system that measures cost per feature shipped, not cost per API call. This reframes the economics in your favor.
• Experiment with dynamic workflows for complex migrations. Krieger's Python-to-TypeScript port over a weekend is a template. Identify one legacy codebase or migration project that has been deferred. Set up a workflow with verification stages and let it run. The ceiling on what is possible is higher than you think.