Anthropic: One Smart Agent with Modular Skills Beats a Thousand Specialists

Redefining Agent Intelligence Through Modular Skills

Barry Zhang and Mahesh Murag from Anthropic presented their research at the AI Engineering Code Summit, and their central finding is striking. Rather than requiring organisations to commission separate agents for finance, legal, compliance, or recruitment, a single robust general agent can be augmented with domain-specific skill packages. These skills are organised collections of files containing instructions, data, and workflows that enable consistent, repeatable task execution.

"We used to think agents in different domains will look very different. The agent underneath is actually more universal than we thought," Zhang explained at the summit.

The analogy to human learning is instructive. A qualified accountant does not become an entirely different person when they move from audit to tax advisory; they draw on a consistent cognitive framework and apply relevant expertise as the task demands. Anthropic argues that well-designed agentic AI systems should operate on exactly the same principle.

For European organisations navigating the EU AI Act's requirements around transparency and human oversight, this architecture has an underappreciated compliance advantage. A single agent with documented, versioned skill modules is considerably easier to audit than a sprawling estate of separately trained specialist models, each with its own data provenance and risk classification to manage.

From Concept to Enterprise Playbook

The skills model addresses well-documented weaknesses in contemporary large language models, particularly their tendency to hallucinate in specialist domains where they lack grounded, task-specific context. Mahesh Murag noted that non-technical professionals in accounting, legal services, and recruitment are already building these skills without requiring deep programming knowledge, packaging their own domain expertise into reusable modules that any compatible agent can deploy.

Within large organisations, skills are effectively becoming internal AI playbooks. A financial reporting skill might contain template structures, IFRS accounting rules, and relevant data source references. A procurement skill might embed supplier approval workflows and EU regulatory obligations. The general agent executes complex tasks by drawing on whichever skill applies, without requiring bespoke model training for each domain.

The comparative economics are worth stating plainly:

• Specialised agents: six to twelve months of development per agent, high ongoing maintenance burden, limited scalability across the organisation.
• Skills-based agent: two to four weeks per skill module, low modular maintenance overhead, high scalability as skills accumulate.
• Traditional software: twelve to twenty-four months, very high maintenance, very limited flexibility.

For European IT leaders facing pressure to demonstrate AI return on investment within a single budget cycle, the skills-based model offers a credible path to faster deployment and lower total cost of ownership.

European Voices: Scepticism and Validation

Not everyone is persuaded that the agent paradigm, whether skills-based or specialist, is as mature as vendor enthusiasm suggests. Lena Meissner, a senior AI policy analyst at the Berlin-based think tank Stiftung Neue Verantwortung, has consistently cautioned that enterprise AI procurement decisions are outpacing the governance frameworks available to manage them. Her view is that modular skill architectures are promising precisely because they create cleaner accountability boundaries, but only if organisations invest in rigorous documentation and testing of each skill module rather than treating them as black-box add-ons.

At the research level, ETH Zurich's AI Centre has been examining the architectural trade-offs between general and specialist agents as part of its work on robust AI systems. Researchers there have noted that generalisation in language models is advancing faster than many anticipated, lending empirical support to the core claim that a single well-designed agent can cover broader ground than previously assumed, provided the surrounding skill infrastructure is well engineered.

The broader industry debate is also shaped by legitimate concerns about marketing inflation. Some vendors are packaging standard language model interfaces as "agents" to command premium pricing, with little genuine agentic capability underneath. European procurement teams should insist on concrete task-completion benchmarks, integration complexity assessments, and documented escalation paths before signing enterprise contracts.

Investment Signals and What They Mean for Europe

Anthropic closed a $30 billion Series G funding round in February 2026, co-led by significant institutional investors and including an $8 billion commitment from Amazon Web Services, which hosts Anthropic's workloads. That level of capitalisation matters for European customers evaluating long-term vendor stability, particularly given the multi-year timescales of enterprise AI transformation programmes.

European cloud and AI infrastructure policy is also relevant here. AWS's investment in dedicated data centre capacity for Anthropic's workloads raises questions about where European enterprise data processed through Claude ultimately resides, and whether that aligns with GDPR obligations and emerging EU AI Act data governance requirements. Organisations in regulated sectors, particularly financial services and healthcare, will need clear contractual assurances on data residency before scaling skills-based deployments.

OpenAI chief executive Sam Altman has separately argued that AI agents are already performing work equivalent to junior employees, while Microsoft's leadership has suggested the technology could flatten corporate hierarchies by reducing managerial layers. Anthropic's skills-based framing is notably more pragmatic than these sweeping structural predictions; it focuses on solving real engineering problems, namely maintenance complexity, development speed, and knowledge transfer, rather than promising organisational transformation on a five-year horizon.