Claude · Published 2026-05-20

Building AI-native at enterprise scale: monday.com, Doctolib, and Delivery Hero

privateReview: unreviewedConfidence: high
enterpriseai-nativeagentscodebaseautomationmulti-agent-reviewmodel-migrationapi-first
Open on YouTube ↗

Summary

Overview

  • Speaker: Rebecca Harbeck, Ruslan Semenov, Alex Kaluzny, Rodrigue Schäfer
  • Channel: Claude
  • Main topic: Integrating AI agents into enterprise-scale legacy codebases
  • Purpose: To share real-world enterprise experiences, architectural shifts, adoption strategies, and lessons learned when deploying AI agents and LLMs at scale. A panel discussion exploring how established tech companies—monday.com, Doctolib, and Delivery Hero—founded between 2011 and 2013 transitioned into AI-native enterprises using Claude and AI-powered coding and workflow tools.

Topic Map

Enterprise Introductions and AI Adoption Context

  • Explanation: Panelists introduce their companies and describe how Claude is powering new missions from workflow management to healthcare and local delivery.
  • Key claims:
    • Companies founded before the LLM era must reinvent themselves as AI-native.
    • AI is shifting platforms from workflow managers to execution platforms.
  • Examples:
    • monday.com using Claude for new AI missions
    • Doctolib building health companions and doctor tools
    • Delivery Hero building the Heron agent for autonomous software delivery
  • Terminology:
    • AI-native
    • Monolith
    • Distributed systems
    • LLM era
  • Why it matters: Establishes the baseline challenges of legacy codebases adopting cutting-edge AI.

Scaling AI Adoption Across Organizations

  • Explanation: Discussion on how to drive adoption not just among engineers, but across product managers, designers, and general staff using internal channels and marketplaces.
  • Key claims:
    • Best engineers find the most creative ways to use new technology.
    • Internal community building accelerates organizational learning.
  • Examples:
    • Doctolib's 'Build with AI' company-wide channel
    • Skills marketplaces and experimental plugin environments
  • Terminology:
    • Skills marketplace
    • Adoption curve
    • Build with AI
  • Why it matters: Cultural adoption is as critical as technical integration for enterprise AI success.

Model Updates and Architectural Choices

  • Explanation: How new model releases impact internal systems, requiring careful evaluation, verification, and migration strategies.
  • Key claims:
    • New model releases require treating upgrades as major system migrations.
    • API-first architecture and unified context layers significantly ease AI integration.
  • Examples:
    • Transitioning between Opus model versions at Delivery Hero
    • Monday.com's migration from 4.5 to 4.6 prompting techniques
  • Terminology:
    • Model versioning
    • A/B testing
    • API-first approach
    • Prompt engineering
  • Why it matters: Managing model shifts prevents systemic regression and maintains high code quality.

Key Points

Involving non-engineers in AI tooling

  • Explanation: Expanding AI tools beyond developers to product managers, designers, and operators multiplies organizational leverage.
  • Evidence: Doctolib reported a large proportion of non-tech staff using Claude for daily work.
  • Practical implication: Companies should provide accessible, sandboxed AI environments for all departments.

The 'Council of Agents' pattern

  • Explanation: Using multiple distinct LLM models to review the same codebase to eliminate blind spots and biases.
  • Evidence: Delivery Hero achieved up to 85% PR success rates using multi-model agent reviews.
  • Practical implication: Do not rely on a single model instance for critical autonomous code generation.

Shifting from generalists to specialists

  • Explanation: Orchestrating specialized AI agents supervised by generalists yields better software delivery outcomes.
  • Evidence: Panel consensus favored orchestrated specialist agents over single generalist coding agents.
  • Practical implication: Design modular agent workflows for complex enterprise pipelines.

Frameworks, Models & Processes

Council of Agents

  • How it works: Multiple different AI models independently review the same code changes and pull requests to detect blind spots, biases, and errors.
  • Components:
    • Primary coding agent
    • Reviewer model instance A
    • Reviewer model instance B
    • Automated merge gateway
  • When to use: When scaling autonomous software delivery and pull request generation in production environments.

Examples & Case Studies

Delivery Hero built 'Heron', an autonomous software delivery agent taking Jira tickets and GitHub issues to production-ready PRs.

  • Illustrates: Autonomous software delivery at scale
  • Lesson: Building for the next model generation rather than the current one ensures longevity.

Actionable Takeaways

  • Immediate:
    • Stop waiting for perfect AI conditions; start experimenting immediately.
    • Integrate AI assistants directly into existing CI/CD and version control workflows.
    • Encourage cross-functional sharing of AI prompts and workflows.
  • Strategic:
    • Invest early in API-first architectures and clean internal boundaries.
    • Establish rigorous multi-model verification ('Council of Agents') for automated code changes.
    • Treat AI model updates as major architectural version upgrades.
  • Questions to investigate:
    • How do we maintain granular security and permission models with autonomous agents?
    • What are the long-term cost tradeoffs of running multi-agent review councils?
    • How will AI change the definition of principal engineering roles?

Claims Worth Verifying

  • Delivery Hero achieved up to 85% success rate on merged pull requests using AI agents. (Metric)
  • Doctolib achieved nearly 100% adoption of building with Claude across product and engineering. (Adoption Statistic)

Notable Quotes

"We are the company which tries to reinvent everything about how we do and what we build for our customers." (at 1:53) "Stop waiting for perfect conditions, perfect use cases, enough AI ready work." (at 46:38)

Compressed Summary

  • Legacy enterprises successfully transitioned to AI-native workflows using Claude.
  • Autonomous agents like Delivery Hero's Heron handle Jira-to-PR pipelines.
  • Multi-model verification ('Council of Agents') drastically increases code merge success rates.
  • Cultural adoption across non-technical roles drives organization-wide AI leverage.
  • Keywords: enterprise, ai-native, agents, codebase, automation
  • Core insight: Enterprise AI adoption succeeds when organizations stop waiting for perfection, embrace multi-agent verification, and integrate AI natively into existing developer workflows.

Core insights

6
Architecturehigh noveltymoderate evidence

For autonomous code generation in production, verification should be delegated to a 'Council of Agents': multiple distinct LLM models independently review the same code changes and PRs to surface blind spots and biases that any single model instance misses. Delivery Hero reports this multi-model review pattern yields up to 85% PR success rates.

Why it matters

It relocates the trust boundary: instead of trusting one coding model or one human reviewer, reliability is produced by model diversity plus an automated merge gateway. This is a concrete design for a verification harness around autonomous agents rather than a slogan about 'evaluation'.

Generalization

Any high-stakes autonomous output pipeline can substitute heterogeneous-model consensus for single-model self-review; diversity across model families acts as an error-correction mechanism analogous to N-version programming.

Multiple different AI models independently review the same code changes and pull requests to detect blind spots, biases, and errors.
Open source video
Do not rely on a single model instance for critical autonomous code generation.
Open source video
Delivery Hero achieved up to 85% PR success rates using multi-model agent reviews.
Open source video
Mental Modelmedium noveltymoderate evidence

Panel consensus favors orchestrating specialized AI agents supervised by human generalists over deploying a single generalist coding agent; the generalist role shifts from writing code to supervising and routing work across modular specialist agents.

Why it matters

It reverses the 'one big agent' default and argues for modular agent workflows as the unit of design in complex enterprise pipelines, with supervision (not generation) as the human contribution.

Generalization

Where task domains are heterogeneous and correctness matters, decomposition into narrow agents with an orchestration/supervision layer outperforms a monolithic generalist agent.

Panel consensus favored orchestrated specialist agents over single generalist coding agents.
Open source video
Orchestrating specialized AI agents supervised by generalists yields better software delivery outcomes.
Open source video
Practicemedium noveltymoderate evidence

Model version releases must be treated as major system migrations, not routine dependency bumps — requiring evaluation, verification, A/B testing, and prompt-technique migration (e.g., monday.com moving from 4.5 to 4.6 prompting techniques, Delivery Hero transitioning between Opus versions).

Why it matters

It makes model identity a versioned, owned architectural surface. Silent model drift or an unmanaged upgrade can regress an entire autonomous pipeline that was tuned to the previous model's behavior.

Generalization

Any system whose behavior is co-adapted with a third-party model should version that model as a first-class dependency with migration playbooks, regression evals, and staged rollout.

New model releases require treating upgrades as major system migrations.
Open source video
Treat AI model updates as major architectural version upgrades.
Open source video
Monday.com's migration from 4.5 to 4.6 prompting techniques
Open source video
Architecturemedium noveltymoderate evidence

API-first architecture plus a unified context layer is reported as the main architectural enabler that makes AI/agent integration tractable in legacy enterprise codebases; clean internal boundaries are where agents plug in.

Why it matters

It claims agents are cheap to add only when the underlying system already exposes stable programmatic interfaces and a shared context surface — the integration cost lives in the legacy boundaries, not the model.

Generalization

Agent readiness is largely a function of existing interface hygiene; retrofitting agents into tightly coupled monoliths is a boundary-refactoring problem before it is a prompting problem.

API-first architecture and unified context layers significantly ease AI integration.
Open source video
Invest early in API-first architectures and clean internal boundaries.
Open source video
Empirical Resultmedium noveltyweak evidence

Adoption by non-engineers is treated as a first-order lever, not a side effect: Doctolib ran a company-wide 'Build with AI' channel and skills marketplaces, with a large share of non-tech staff using Claude daily and near-total adoption across product and engineering.

Why it matters

It implies the AI-native organization is designed for distributed authorship of prompts and workflows, which changes internal tooling, sharing, and permission requirements rather than remaining a developer-only surface.

Generalization

Organizational leverage from AI scales with the breadth of contributors; sandboxed, shareable prompt/workflow environments are infrastructure, not perks.

Doctolib reported a large proportion of non-tech staff using Claude for daily work.
Open source video
Doctolib achieved nearly 100% adoption of building with Claude across product and engineering.
Open source video
Companies should provide accessible, sandboxed AI environments for all departments.
Open source video
Mechanismmedium noveltymoderate evidence

Delivery Hero's Heron is an autonomous software delivery agent that turns Jira tickets and GitHub issues into production-ready PRs, and the stated design principle is to build for the next model generation rather than the current one.

Why it matters

It shows an end-to-end ticket-to-PR pipeline as a concrete agent runtime target, and introduces forward-compatibility with stronger future models as an explicit design constraint against tuning tightly to today's model limits.

Generalization

Systems that abstract away the assumption 'the model cannot do X yet' avoid being rewritten each model cycle; capability assumptions should be parameters, not hard architecture.

Delivery Hero built 'Heron', an autonomous software delivery agent taking Jira tickets and GitHub issues to production-ready PRs.
Open source video
Building for the next model generation rather than the current one ensures longevity.
Open source video

Deep dives

5

Cost and failure-mode economics of multi-model review councils

Research question

At matched defect-detection quality, how do cost-per-merged-PR, latency, and reviewer-failure correlation scale with the number of heterogeneous reviewer models in an autonomous code-review council?

Why

The panel reports a headline success rate (up to 85% PR success) but explicitly leaves the long-term cost tradeoff unresolved. Without cost-per-merged-PR curves, 'Council of Agents' stays an unfalsifiable reliability story rather than an engineering tradeoff that can be budgeted and scheduled inside a delivery pipeline.

Multiple different AI models independently review the same code changes and pull requests to detect blind spots, biases, and errors.
Open source video
Delivery Hero achieved up to 85% PR success rates using multi-model agent reviews.
Open source video
What are the long-term cost tradeoffs of running multi-agent review councils?
Open source video
Source video

Model version identity as a first-class architectural dependency

Research question

How much measured regression occurs when a tuned agent pipeline upgrades its underlying model without re-running evaluation and migrating prompt techniques, and what migration playbook minimizes that regression?

Why

If model releases are migrations, then model identity must be pinned, owned, and regression-tested like a framework major version. This determines whether an enterprise needs a standing model-migration function with staged rollout, A/B testing, and prompt-technique rework every model cycle.

New model releases require treating upgrades as major system migrations.
Open source video
Monday.com's migration from 4.5 to 4.6 prompting techniques
Open source video
Managing model shifts prevents systemic regression and maintains high code quality.
Open source video
Source video

Interface hygiene as the real cost of agent integration in legacy monoliths

Research question

Can agent task success in a legacy enterprise codebase be predicted from a measurable score of interface cleanliness and unified-context-layer coverage, independent of which model is used?

Why

The claim relocates AI integration cost out of prompting and into boundary refactoring. If agent readiness is a function of pre-existing interface hygiene, then agent programs in pre-LLM monoliths should budget for API-first refactoring and a shared context layer before, not after, model selection.

API-first architecture and unified context layers significantly ease AI integration.
Open source video
Invest early in API-first architectures and clean internal boundaries.
Open source video
Source video

Least-privilege agent identity and auditable action provenance

Research question

Can capability-scoped agent identities plus an automated merge gateway satisfy enterprise permission requirements for autonomous ticket-to-PR agents without materially reducing autonomous throughput?

Why

Broad non-engineer authorship and autonomous agents compound the permission surface simultaneously: more humans issuing instructions while agents act across many system boundaries. Existing permission models were not designed for principals that both write code and open pull requests unsupervised.

How do we maintain granular security and permission models with autonomous agents?
Open source video
Expanding AI tools beyond developers to product managers, designers, and operators multiplies organizational leverage.
Open source video
Companies should provide accessible, sandboxed AI environments for all departments.
Open source video
Source video

Capability-parameterized agent runtimes built for the next model generation

Research question

Do runtimes that express 'the model cannot do X yet' as declared capability parameters, rather than hard-coded fallbacks, require measurably less rework across two successive model releases?

Why

Every hard-coded compensation for a current model limitation is technical debt that expires on the next release. Making capability assumptions configuration converts model progress from a rewrite trigger into a parameter change, which is the concrete mechanism behind the 'build for the next generation' principle.

Building for the next model generation rather than the current one ensures longevity.
Open source video
Delivery Hero built 'Heron', an autonomous software delivery agent taking Jira tickets and GitHub issues to production-ready PRs.
Open source video
Source video

Article ideas

4

Your Next Model Upgrade Is a Migration, Not a Dependency Bump

Model identity is now a versioned architectural surface: teams that treat a model release as a routine dependency bump will silently regress an entire autonomous pipeline, while teams that run model upgrades as staged migrations with regression evals and prompt-technique rework preserve code quality across cycles.

Angle

Turn the monday.com 4.5-to-4.6 prompting migration and the Delivery Hero Opus transitions into a repeatable migration playbook: pin, eval, A/B, migrate prompts, roll out, then re-evaluate the reviewer pool.

Source video

The Council of Agents Is N-Version Programming With a Bill Attached

Heterogeneous multi-model review is a genuine error-correction mechanism in the tradition of N-version programming, but it is only defensible engineering once you publish cost-per-merged-PR, latency overhead, and reviewer-failure correlation instead of a single success-rate headline.

Angle

Argue that model diversity moves the trust boundary from a single model to an ensemble/consensus merge gateway, then demand the economics that the panel left unresolved.

Source video

Agents Don't Fix Dirty Boundaries: Agent Readiness Is Interface Hygiene

The cost of making a legacy enterprise codebase AI-native is dominated by API-first boundary refactoring and building a unified context layer, not by prompt engineering, so agent programs should be sequenced as platform work rather than as model procurement.

Angle

Reframe agent adoption as a boundary-and-context engineering program, with agent task success as the dependent variable and interface cleanliness as the independent one.

Source video

Everyone Authors Prompts Now, and the Permission Model Is the Adoption Blocker

Broad non-engineer adoption multiplies organizational leverage only when sandboxed, shareable AI environments exist, and the moment those same contributors direct autonomous agents the permission model, not model capability, becomes the binding constraint.

Angle

Connect Doctolib's company-wide Build with AI channel and skills marketplaces to the unresolved question of granular permissions for autonomous agents.

Source video

Project ideas

4

Council Bench

beyond-evals

Adding heterogeneous reviewer models reduces missed-defect rate per review-dollar up to a knee, beyond which reviewer failures become correlated and marginal cost exceeds marginal defect reduction.

Proof of concept

Run a primary coding agent plus N reviewer models drawn from different model families (N = 1..4) over a held-out set of real pull requests with known post-merge defects, gating merges on reviewer consensus, and log cost, latency, and escaped defects per configuration.

Measurement

Defect escape rate versus cost-per-merged-PR curve; reviewer pairwise failure-correlation matrix; p95 review latency per additional reviewer.

Source video

Model Migration Harness

beyond-evals

Pipelines that pin model versions and re-run a fixed evaluation suite plus prompt-technique migration per release show a materially smaller behavioral regression than pipelines that upgrade in place.

Proof of concept

Replay a fixed suite of production agent tasks against an old and a new model version, both with and without prompt re-migration, and diff outputs, eval scores, and human re-review rates.

Measurement

Eval pass-rate delta and output-divergence rate between model versions; share of tasks requiring human re-review after unmanaged versus managed upgrade.

Source video

Least-Privilege Agent Passport

gatehouse

Capability-scoped, short-lived agent identities with signed action provenance block unauthorized actions under adversarial prompt injection while costing less than a defined throughput threshold versus an unscoped agent.

Proof of concept

Instrument a ticket-to-PR agent with per-repository and per-path scoped credentials and an append-only provenance ledger of every tool call, then run a prompt-injection and privilege-escalation suite against it.

Measurement

Unauthorized action attempts blocked versus attempted; autonomous PRs per day with and without scoping; time to reconstruct a full audit trail for a single change.

Source video

Capability-Parameterized Agent Runtime

new

Expressing model-limitation assumptions as declared capability parameters with fallbacks, rather than hard-coded workarounds, reduces rework hours per model release by a measurable margin across two successive generations.

Proof of concept

Build a Heron-style ticket-to-PR pipeline in which every 'the model cannot do X yet' assumption is a named capability parameter with a configurable fallback, then upgrade the underlying model and record what must change.

Measurement

Lines of code changed and engineering hours of rework per model release; task success-rate delta after upgrade; count of capability parameters flipped without code edits.

Source video

Architectural implications

4

Multi-model code review councils add a model-diversity review tier inside the delivery pipeline.

Before

Code generation and review rely on a single model instance plus human review.

After

A primary coding agent feeds reviewer model instances A and B, whose outputs gate an automated merge gateway.

Consequence

The merge decision becomes an ensemble/consensus function, raising reliability but adding cost, latency, and a new place where model coordination can fail.

Source video

Legacy monoliths are refactored toward API-first boundaries and a unified context layer to host agents.

Before

AI tooling is bolted onto existing systems ad hoc.

After

Stable programmatic interfaces and a shared context layer become the designated integration surface for agents and assistants.

Consequence

Agent integration effort shifts from prompt work into boundary and context-layer engineering.

Source video

Model version identity becomes an owned, versioned architectural dependency.

Before

Model usage is an unpinned implementation detail; upgrades are routine.

After

Model upgrades are executed as major migrations with evaluation, A/B testing, and prompt-technique changes.

Consequence

Organizations need migration playbooks and regression evaluation for models, similar to framework upgrades.

Source video

Human engineering roles re-center on supervision and orchestration.

Before

Generalist engineers author code directly.

After

Generalists supervise and orchestrate networks of specialized agents that do the authoring.

Consequence

Role definitions, review gates, and escalation paths must be redesigned around supervision rather than authorship.

Source video

Tradeoffs and failure modes

4

Multi-model review councils for autonomous code changes

Benefit

Higher PR success rates and reduction of single-model blind spots and biases via independent heterogeneous review.

Cost or risk

Long-term cost tradeoffs of running multi-agent review councils are explicitly unresolved, implying added model/latency/infra spend per change.

What are the long-term cost tradeoffs of running multi-agent review councils?
Open source video
Source video

Single-model autonomous code generation

Benefit

Simple, low-coordination pipeline with one agent doing generation.

Cost or risk

Concentrated blind spots and biases; explicitly warned against for critical autonomous code generation.

Do not rely on a single model instance for critical autonomous code generation.
Open source video
Source video

Treating model upgrades as system migrations

Benefit

Prevents systemic regression and maintains code quality across model releases.

Cost or risk

Material recurring migration burden each model cycle, including re-validating prompts and re-running evaluation.

New model releases require treating upgrades as major system migrations.
Open source video
Source video

Broad non-engineer access to AI tooling

Benefit

Multiplies organizational leverage by expanding AI usage beyond developers.

Cost or risk

Raises unsolved questions about granular security and permission models, especially once agents act autonomously.

How do we maintain granular security and permission models with autonomous agents?
Open source video
Source video

Open questions

4

How do we maintain granular security and permission models once agents act autonomously inside enterprise systems?

Why unresolved

The summary lists this as a question to investigate; autonomous agents spanning ticket-to-PR pipelines cross many permission boundaries that existing models do not cover.

Research direction

Design and benchmark capability-scoped agent identities, least-privilege tool grants, and auditable action provenance for autonomous code and delivery agents.

Source video

What are the long-term cost tradeoffs of running multi-agent review councils in production?

Why unresolved

Only a success-rate benefit (up to 85% PR success) is reported; no cost, latency, or token economics are given.

Research direction

Measure cost-per-merged-PR and latency overhead of N-reviewer councils versus single-model review at matched quality, including diminishing returns as reviewer count grows.

Source video

How will AI change the definition of principal engineering roles?

Why unresolved

The panel favors generalist supervision of specialist agents but does not define the resulting seniority ladder or skill set.

Research direction

Study how supervision, orchestration design, and evaluation ownership map onto existing senior/staff/principal scopes.

Source video

How should systems be architected so they improve, rather than get rewritten, as stronger models arrive?

Why unresolved

The stated principle is to build for the next model generation, but no concrete abstraction is described for doing so.

Research direction

Prototype capability-parameterized agent runtimes where model-dependent assumptions are configurable and can be re-evaluated per model release.

Source video

Key claims

7
factualVerification needed

Delivery Hero achieved up to 85% success rate on merged pull requests using AI agents.

Evidence

Delivery Hero achieved up to 85% success rate on merged pull requests using AI agents. (Metric)

Question

What was the denominator (all PRs vs. agent-authored PRs), the definition of 'success', and the human-review/merge criteria behind the 85% figure?

Source video
factualVerification needed

Doctolib achieved nearly 100% adoption of building with Claude across product and engineering.

Evidence

Doctolib achieved nearly 100% adoption of building with Claude across product and engineering. (Adoption Statistic)

Question

How was adoption measured (active usage, licensed seats, or merged AI-assisted work), and over what time window?

Source video
causalVerification needed

Using multiple distinct models to review the same code changes removes blind spots and biases present in single-model review.

Evidence

Multiple different AI models independently review the same code changes and pull requests to detect blind spots, biases, and errors.

Question

Does model diversity measurably reduce missed-defect rates versus a single model with equivalent review budget, and do review councils exhibit correlated failures?

Source video
comparativeVerification needed

Orchestrated specialist agents supervised by generalists outperform a single generalist coding agent for software delivery.

Evidence

Panel consensus favored orchestrated specialist agents over single generalist coding agents.

Question

Under what task mix and orchestration overhead does specialization beat a single strong generalist agent on delivery outcomes?

Source video
causalVerification needed

New model releases must be handled as major system migrations rather than routine upgrades.

Evidence

New model releases require treating upgrades as major system migrations.

Question

How much measured regression occurs when a tuned pipeline upgrades models without re-evaluation or prompt migration?

Source video
causalVerification needed

API-first architecture and unified context layers significantly ease AI integration in legacy enterprises.

Evidence

API-first architecture and unified context layers significantly ease AI integration.

Question

Can integration effort (or agent task success) be quantified against a measure of interface cleanliness and context-layer coverage?

Source video
predictionVerification needed

Building for the next model generation rather than the current one ensures longevity of agent systems.

Evidence

Building for the next model generation rather than the current one ensures longevity.

Question

Which concrete abstractions avoid model-specific assumptions, and do systems designed this way require measurably less rework across model releases?

Source video

Connections

4