Mike Krieger

Mike Krieger co-founded Instagram, where he was chief technology officer, and later co-founded the news-discovery app Artifact. At his AI Engineer World’s Fair 2026 appearance, he led Anthropic Labs, following his work as Anthropic’s chief product officer. His career spans interface design, the infrastructure behind widely used consumer products, and experiments in giving AI agents substantial responsibility for engineering work.

From interface design to Instagram’s infrastructure

Krieger grew up in São Paulo, Brazil, and studied Symbolic Systems at Stanford, combining computing, design, and cognitive science. He pursued human-interaction coursework and met Kevin Systrom through the Mayfield Fellows entrepreneurship program. Before Instagram, he worked in user experience and front-end development at Meebo. Building a service that people could use intuitively was his starting point; keeping it running under unexpected demand became his next education.

Krieger and Systrom founded Instagram in 2010 after narrowing their earlier check-in product, Burbn, around photo sharing. Its early growth pushed Krieger into cloud infrastructure and systems administration, with help from people who had deeper operational experience. Do the simple thing first began as a necessity for a two-person development team and remained an engineering principle as Instagram grew.

As co-founder and CTO, Krieger helped scale Instagram to more than a billion users and grow its engineering organization to over 450 people. He also stayed involved in detailed technical problems. In late 2016, he made wide-color photography a personal focus, helping the team preserve the richer colors captured by newer iPhones. Capture, resizing, filtering, export, and delivery each offered opportunities to lose that information. The team used a diagnostic image whose Instagram logo disappeared when wide-color data was discarded, making a subtle pipeline failure immediately visible.

Instagram’s early scaling problems also shaped his approach to operations. He learned to measure likely failure signals before an outage, when engineers would need a baseline to distinguish normal behavior from trouble. Feature flags and dynamic configuration let the team adjust rollouts and respond to load within seconds. In his 2026 discussion of engineering with agents, he connected those practices to AI development: faster experimentation still needs production measurements and ways to control changes.

Public data and AI-assisted news

In 2020, Krieger worked on rt.live’s daily COVID-19 estimates. Its automated pipeline ran regional models in parallel, retained dated data snapshots, checked convergence, and flagged unexpected changes before publication. When source data needed correction, affected regions could be rerun individually. The work brought reproducibility and operational checks to a public data product whose results had to remain understandable and up to date.

Krieger subsequently reunited with Systrom to build Artifact, serving as co-founder and CTO. The app personalized news recommendations and explored how language models could improve reading itself. Its clickbait headline rewriting feature generated more factual titles from article content. Early tests showed that supplying the original headline encouraged the model to reproduce its framing, so the team excluded it from the input. They tested examples and cleaned up unwanted output, while animation and consistent AI iconography explained the wait and identified the generated result. Rewrites were initially private to the requesting reader; replacing a headline for everyone required human review.

Returning to hands-on building at Anthropic

Yahoo acquired Artifact in March 2024 to integrate its personalization technology and features into Yahoo products. Krieger joined Anthropic as chief product officer that May, overseeing product engineering, product management, and design. In January 2026, he moved into Labs to build alongside Ben Mann, while Ami Vora took over the product organization. He explained at World’s Fair that increasingly capable models had made his weekend building experiments compelling enough to pursue hands-on creation as his daily work.

That transition changed how Krieger delegated engineering. Rather than breaking an idea into implementation tasks himself, he increasingly described the desired end state and asked Claude to work toward it, raising questions and explaining tradeoffs along the way. One experiment was a weekend migration of a substantial internal Python project to TypeScript for a better deployment path. He set up a workflow that repeatedly verified the result and compared the implementations, returning on Monday to what he described as a completed, deployable port.

His enthusiasm for ambitious migrations comes with an operational qualification: find a place to begin incrementally. At Instagram, the team’s MonkeyType work captured types actually used in production and mapped them back into the Python codebase. Krieger sees similar value in using production observations and segmented tests to inform language-model-assisted conversions, rather than relying only on generated code or attempting to replace everything overnight.

Understanding changes and expanding agent responsibility

• Human comprehension: Faster code generation also makes human comprehension a constraint. Krieger described architectural reviews in which a large pull request was difficult to assess even when the code appeared plausible. His team began sharing explanations of a change’s intent and tradeoffs through Claude Code artifacts. He uses Claude to investigate his questions about an implementation while keeping consequential reviews human-directed. The aim is to understand what the system will do, why particular choices were made, and what production measurements should confirm.
• Ongoing agent responsibility: Shared agent workflows have expanded his sense of what delegation can include. Watching a colleague assign an agent ongoing responsibility for a part of the codebase—monitoring feedback, taking on fixes, and responding to API changes—revealed possibilities beyond an isolated coding session. These asynchronous workflows make examples visible to other team members and let agents retain context across tasks. Krieger also argues for giving agents enough tools to recover from failure: if a built-in PDF parser cannot handle a document, an agent may still succeed by writing and running a script.

Two-week bets and connected products

Within Labs, two-week product bets provide a structure for rapid experimentation. Every project faces a decision to continue, change direction, or stop. Temporary groups draw participants from product and engineering, with a responsible lead who usually does not manage the other members. Separating project assignments from management relationships lets a group disband without reorganizing the company every two weeks. Engineering managers still provide coaching and development and help people find suitable work. Experiments that gain traction can acquire dedicated teams; Krieger described Claude Design as one product that moved from an ad hoc group toward a more established organization.

He sees unnecessary product divisions as an obstacle to useful delegation. Users should not have to carry context manually between design, coding, chat, and other workspaces. His example was a user who had worked out what to build in one session, then had to ask for a paragraph to paste into Claude Code. Better integration would let the tools pass work to one another. He also favors replacing older AI features when newer capabilities serve the same need better, rather than allowing each generation to add more product complexity.

Startup opportunities as coding gets easier

Krieger remains optimistic about startups that understand a particular industry or group of users deeply. Easier coding accelerates experiments, but choosing a valuable problem, reaching customers, and learning from them remain demanding. Small teams can concentrate on needs that a larger platform, with its existing products and integrations, may struggle to address as closely. He also acknowledges that some functions may become model capabilities or skills rather than support a separate product.

Financial applications make the tradeoff especially concrete. Krieger wants agents to create analyses, dashboards, and workflows flexibly while drawing on data with dependable verification, provenance, and audit history. Entirely free-form work can create confusion; systems built for auditability can also be too rigid for agent workloads. He sees opportunity in combining trustworthy data foundations with room to build new applications above them.

Making rapid experimentation sustainable

His approach to leadership includes making the pace sustainable. Krieger encourages genuine time offline and cautions against tying personal success to every model launch or competitive reversal. When a Labs experiment ended, he told colleagues that he was sad and frustrated it had not worked. Acknowledging his own disappointment made room for others to express theirs before deciding what to do next. Rapidly ending experiments requires care for the people who invested in them.