Anthropic Says AI Agents Are Shifting Software From Tools to Workplace ‘Colleagues’

Software is shifting from a tool people operate into a workplace ‘colleague’ that can execute tasks end-to-end—an evolution that could push AI adoption beyond developers and into the broader ranks of ‘knowledge workers’, according to Anthropic’s applied AI team.

Speaking Thursday ET (Wednesday UTC) at MetaCon 2026, a major AI conference co-hosted by TokenPost in Seoul, Jang Dongjin, Applied AI Architect at Anthropic, outlined how AI agents are reshaping day-to-day work inside organizations and what companies should prioritize as they pursue ‘AX’ (AI transformation). His session, titled “Agent Evolution: AI Has Started Working”, centered on the idea that the decisive variable in AI transformation is not only model capability, but the redesign of human roles around delegation and verification.

Jang argued that the relationship between people and software has changed fundamentally over the past two decades. From the 2000s through the early 2020s, he said, software primarily boosted performance and quality as a ‘tool’. Now, as agentic systems mature, users increasingly share goals rather than issue step-by-step instructions—and expect completed outputs rather than isolated answers—leading software to be perceived as a ‘colleague’ that can collaborate on outcomes.

At the heart of that shift is the rise of AI agents structured around a repeating cycle of ‘planning–action–reflection’. In Jang’s framing, this loop enables systems to set sub-goals, execute tasks, validate results, and iterate when shortcomings are detected—moving large language models beyond one-shot Q&A into continuous, work-oriented execution.

He pointed to code assistants as a clear example of how role redefinition can unlock measurable gains. Early coding copilots typically suggested snippets, leaving engineers to integrate and complete the work. With the emergence of tools such as ‘Claude Code’, he said, the workflow increasingly flips: AI generates large portions of code while humans set direction, define constraints, and review output for correctness and intent. That pattern, he added, is spreading from individual productivity improvements to team-level throughput as organizations change how work is assigned and validated.

Jang cited global companies including Spotify as examples where software engineering roles are increasingly oriented toward reviewing and steering AI-produced output rather than writing every line manually. “The most important success factor in AI transformation is changing the user’s role,” he said, arguing that adoption accelerates when execution is delegated to AI and humans focus on judgment—setting objectives, assessing quality, and deciding what ships.

In his view, the next major expansion is not another developer tool, but agent adoption across the everyday responsibilities of office work. He emphasized that the ‘next target after developers is knowledge workers’, describing a future in which agents handle routine but complex tasks such as meeting-note summarization, action-item extraction, document drafting, and prototyping—triggered primarily through natural language prompts.

During the session, Jang demonstrated agent-driven workflows including a desktop agent dubbed ‘Claude Core’ designed to carry out multi-step tasks, and a design canvas called ‘Claude Design’ that can generate prototypes through conversational instructions. While the demos served as proof points, Jang stressed that the key story is the broadening scope of labor that can be delegated—not any single product feature.

For enterprises mapping out an ‘AX’ roadmap, Jang recommended moving in stages: start with individual productivity, expand into process automation, and then push toward product and service innovation. He laid out three principles for implementation: focus first on problems with clear objectives and measurable evaluation criteria; build ‘context’ and connectors so AI systems can securely use internal data and institutional knowledge; and delegate work aggressively to AI while establishing verification frameworks to manage risk and quality.

MetaCon 2026 runs July 3–4 in Seoul at COEX Grand Ballroom under the theme “AI Makers Rise,” bringing together companies and builders to share strategies and execution experience spanning enterprise innovation, marketing, and investment. Jang’s remarks underscored a growing industry consensus: as agentic systems become more reliable, competitive advantage may hinge less on introducing AI and more on redesigning workflows so humans and machines can co-produce outcomes at scale.