CHAPTERS
Panel setup: pivoting pre-LLM companies to AI-native enterprise
Rebecca (Anthropic GTM) frames the discussion: three established companies—founded before LLMs—are now transforming into AI-native enterprises. The focus is on how they retrofit AI into mature products, orgs, and codebases rather than starting greenfield.
Where Claude is already mission-critical at each company (and what the codebase looks like)
Each panelist introduces their role and the Claude-powered systems that matter most. They highlight the reality of large, imperfect legacy codebases—often monolithic—and the ongoing transition toward more modular architectures.
Delivery Hero’s HeroGen: autonomous software delivery from ticket to PR
Delivery Hero describes building an agent that takes a Jira ticket or GitHub issue and produces a production-ready pull request. They share adoption metrics that suggest strong and accelerating usage across teams and subsidiaries.
Doctolib’s approach: scale adoption by empowering the whole engineering org
Doctolib focuses less on one central AI team and more on enabling every engineer (and beyond) to innovate with Claude. Platform teams concentrate on enablement, best practices, and scaling what works—rather than owning all AI usage directly.
Doctolib’s “skills marketplace” and developer environment for reusable AI workflows
Doctolib operationalizes AI usage through discoverable, reusable “skills” that teams can share and iterate on. New developers get an environment with tools pre-wired and a pathway to experiment via plugins and community feedback loops.
monday.com’s customer-facing AI: monday Vibe prompt-to-app builder
monday.com highlights monday Vibe, a tool that turns a simple prompt into a detailed PRD and then a working application in minutes. Early success was accelerated by prior investment in an open platform that let Vibe behave like an external developer using the same APIs/SDKs.
When new Claude models ship: what changes operationally inside each org
The panel explains what happens from model release to production rollout, emphasizing excitement, evaluation, and real end-to-end testing. They note that model upgrades can be step-change moments but also require significant re-optimization.
monday.com’s model migration reality: orchestrators, workflows, and re-prompting
monday.com describes Vibe as a multimodal system with an Opus-driven orchestrator and deterministic sub-workflows. A model change can break previously optimized prompts, requiring deep rethinking and renewed prompt engineering—treated like adopting a different system, not a drop-in upgrade.
If starting today: Delivery Hero’s integration-first architecture and ‘council of agents’ review
Delivery Hero prioritizes fitting agents into existing engineering workflows (Jira/GitHub/GitLab) rather than forcing new UIs. They also highlight an architectural boost to quality: multiple models reviewing the same code to reduce blind spots and bias, driving higher acceptance rates.
If starting today: Doctolib on standardization, smaller codebases, and new bottlenecks
Doctolib contrasts adoption ease in newer, opinionated service architectures versus a legacy monolith with many historical patterns. They also observe that as coding accelerates via agents, organizational and process bottlenecks become newly visible and more expensive.
If starting today: monday.com’s API-first, identity, and permissions for agent-first worlds
monday.com emphasizes that agents—internal and external—need reliable API access and clear service boundaries. They also call out the complexity of treating agents as first-class users, requiring rethinking identity and authorization models—especially with granular permissions.
Lightning round: specialist vs generalist agents, how engineer roles are changing, and daily metrics
All three lean toward multiple specialized agents coordinated by a generalist. They describe role shifts: senior engineers and even non-engineers become more hands-on builders, while teams focus on orchestration, context layers, and using failures as discovery signals. Metrics range from PR success rates to reliability and failure analysis.
Advice to teams ‘six months behind’: mandate usage, start with codebase analysis, don’t wait for perfection
The closing guidance is pragmatic: adoption comes from using agents on real work immediately. Delivery Hero suggests setting a goal for every team to ship an end-to-end AI-built feature; Doctolib suggests starting by having an agent analyze the codebase and propose an improvement plan; monday.com urges engineers to stop waiting for ideal refactors and apply AI to repetitive toil now.
