Writing software is the killer use case for LLMs. It has always been a perfect fit – expensive, digitally native, with formal language and explicit patterns. Models have been improving exponentially on coding benchmarks. But it took the stepchange to Anthropic’s Sonnet 3.5 in June 2024 to move us beyond coding “assistants” to “partners”, with its ability to generate thousands of tokens of coherent code in one shot. We’ve come far enough that Andrej Karpathy, co-founder of OpenAI, coined the term “vibe-coding”. He described a shift from hand-coding to prompting AI to write code – and perhaps even forgetting code exists.

Anthropic CEO Dario Amodei has even more of an incentive to predict the end of writing code, but it is no longer a rare opinion. AI code generation tools are becoming indispensable for technical and non-technical users alike. AI code editors like Cursor and Windsurf are becoming table stakes for engineering teams, with increasingly-dramatic influences on how code is getting shipped. We’ve heard of some teams already ripping out Microsoft’s VSCode entirely. AI text-to-app tools like Lovable and Bolt.new are building incredible communities and a Cambrian explosion of experiments. Large incumbents are jumping into this space. Google launched Firebase Studio spanning backends, front ends, and mobile apps & the likes of Canva are building domain-specific tools. OpenAI is rumoured to be in talks to buy Windsurf for $3 billion. The revenue is tracking: Cursor hit $200m of ARR, and GitHub Copilot drove 40% of Github’s revenue growth last year. 

A cursor-y landscape of agencies

Clearly, then, these tools are revolutionising what gets done within companies.

If anyone can build, and fast, what happens to software agencies who build for others? Think of services like: 

• Web agencies – Marketing sites, ecommerce storefronts and CMS platforms;
• Custom app builders – Bespoke software tailored to specific business processes – internal tools, customer portals and complex workflow systems;
• Mobile app specialists – Native and cross-platform apps optimized for mobile UX.
• Platform integrators – ERP, CRM, legacy and third-party integrations into unified workflows;
• Data & infrastructure consultancies – Cloud environment management, data pipelines and analytics platforms;
• QA & testing – Manual and automated software testing.

Just taking the ten largest service providers in this space (the likes of Accenture, TCS, Infosys) amounts to $180 billion in a service revenue opportunity. And this isn’t just legacy IT. Custom software is still the default path for internal tools, customer portals, analytics environments, and regulated workflows. Web and mobile agencies dominate the SMB end of the market, while enterprise spend is concentrated among system integrators, app developers, and cloud/data consultancies. 

The “minimum lovable agency”, today and tomorrow

Agencies sell two things: time and trust. Historically, the inherent complexity of software development justified both. 

But LLMs are unbundling the “time” element in an agency proposition.

• Discover: LLMs swiftly generate proofs-of-concept and draft initial specifications, streamlining initial ideation. 
• Build: LLMs automates CRUD scaffolding, boilerplate code, and increasingly complex workflows with agents. AI tooling can complete tasks in minutes that previously took hours or days. In a (now common) time & materials commercial set-up, this means lost billable hours and lost complexity that drives outsourcing in the first place. 
• Maintain: LLMs can handle regression testing, uptime monitoring, and routine alerts effectively. 

AI tools make it easier than ever to start your own agency, with fewer minimum headcount requirements. Undifferentiated agencies tapping into lower risk and/or lower value use-cases risk a race to the bottom on price. Lovable and Replit, for example, are proactively cultivating networks of preferred agency partners.

Yet, we think there remains a stubborn case for (reinvented) agencies. LLMs today need a partner to untangle each real-world environment. Agencies will continue to own the “first‑mile” work of translating vague needs into precise requirements and the “last‑mile” work of weaving together a coherent bespoke set-up:

• Problems and goals are often fuzzy. Many requirements emerge only after sketched prototypes and stakeholder interviews. The best agencies use early sessions to reveal hidden priorities. Clients frequently remark that the finished product feels nothing like their original brief – and exactly what they needed.
• Each context is incredibly messy, vague and evolving. Each client has unique environments, operations, and constraints. A “great” output here is not clearly defined, unlike in the coding benchmarks. Agencies absorb these systems-level issues and ambiguity, and convert them into maintainable software.
• Clients like outcomes. Trust and reliability matters a lot in production. Clients want polished deliverables, security and guaranteed uptime. Agencies promise this with service level agreements and manage handovers. In some instances, this is an offloading of effort too – not everyone wants to be a developer, for every project. 

The democratisation of software is expanding the overall universe of what gets built, too. If everyone can be a software developer, it means more prototypes, which means more projects will be validated (and likely start-up ideas too!) – a proportion of which agencies will help bring to life. In a possible world where all software is custom, that is a very large universe indeed.

Vibe-coding may commoditise or democratise code (depending on your perspective!). But it cannot yet capture the full messiness that sits around it. Services aren’t dead but they’ll be reinvented – more outcome-driven than ever and the best agencies adopting the new universe of tools into their workflows.
As investors, we are excited by founders tapping into this huge universe of spend – either enabling channel partners and/or competing for the same use-cases. If that’s you, please reach out to shamillah@dawncapital.com