AI Agents: Why Waiting Is No Longer an Option.
//
IT & Management Consulting, IT Strategy
Bei Anthropic schreibt die KI 90 Prozent des Codes für neue Claude-Versionen. OpenAI hat mit GPT-5.2-Codex sein „fortschrittlichstes agentisches Coding-Modell“ veröffentlicht. Salesforce spricht vom Jahr der „Multi-Agentic Enterprise“. Was nach Silicon-Valley-Übertreibung klingt, ist der neue Alltag in führenden Tech-Unternehmen und ein Weckruf für den deutschen Mittelstand und Großunternehmen.
What Sets AI Agents Apart from Existing Tools
The term “AI agent” is used loosely. A straightforward definition: AI agents are systems that autonomously perform tasks across multiple steps. They analyze context, plan courses of action, use external tools, and adapt their approach based on interim results without requiring each step to be initiated manually.
The difference from earlier AI tools is fundamental: ChatGPT or early versions of Copilot respond to individual requests. The current generation of AI agents – whether Claude Code, OpenAI Codex, or GitHub Copilot – takes on complex tasks autonomously: They navigate repositories independently, understand context across thousands of files, carry out multi-step, complex tasks, and validate and document their own results. These systems do not respond to individual commands but work independently toward goals.
The Current Situation
In the recent weeks, software development agents have reached an impressive level of maturity. The latest reports from the tech industry are documented reality. At Anthropic, the proportion of AI-generated code company-wide ranges from 70 to 90 percent. Boris Cherny, Head of Claude Code, explained in late January that 100 percent of his code is generated by AI. Cherny’s approach illustrates the new paradigm: He runs 10 to 15 AI agent sessions in parallel: five in the terminal, five to ten in the browser, plus mobile sessions that he starts in the morning and checks later. The developer becomes an orchestrator who delegates tasks and reviews results.
At the same time, OpenAI has released GPT-5.2 Codex, its most powerful agent-based coding model. It can handle many tasks in parallel and is available to enterprise customers with SOC 2 compliance and single sign-on. The capabilities of the leading platforms are converging, and competition is increasingly taking place in the areas of integration, governance, and user experience. However, this development is not limited to development environments.
Coding was just the beginning
In mid-January, Anthropic launched Claude Cowork, which was developed in just one and a half weeks—largely by Claude Code itself. The product brings agent capabilities to the desktop for non-programmers: document generation, data extraction, and automated workflows.
Claude in Excel demonstrates just how far the development has already come: For example, the add-in analyzes complex multi-sheet workbooks, builds integrated financial models from income statements, balance sheets, and cash flow statements, and performs DCF analyses with sensitivity tables. Every action is documented with cell-level references, allowing users to trace each step.
Salesforce is driving the same trend in the enterprise segment with Agentforce. The platform has long since moved beyond serving only sales and service; it now automates processes in HR, IT, finance, and legal. The Spring ’26 release integrates these modules into a unified architecture. Salesforce refers to this as the transition to the “Agentic Enterprise”—an organization in which human employees and AI agents collaborate as an integrated ecosystem. What does this development mean for specialists and executives?
From Programming to Management
Wharton professor Ethan Mollick observes a fundamental shift: “It’s interesting to see how some of the best-known software developers in the major AI labs report that their work is shifting from primarily programming to primarily managing AI agents.” Software development is the first field in which AI agents have matured and is thus one of the first professions to feel the impact of this shift particularly keenly.
The competitive edge does not come from the tool itself, but from the organizational ability to deploy AI agents safely and effectively. These lessons cannot be bought, and it is nearly impossible to catch up later.
It won’t be the last one. Mollick’s central thesis: The skills needed to work with AI agents are classic management competencies. Anyone who can explain what they need, provide effective feedback, and evaluate work results will be able to work with agents. “The skills that are so often dismissed as ‘soft’ turned out to be the hard ones.” Management has always been based on scarcity: You delegate because you can’t do everything yourself and because talent is limited and expensive. AI is changing this equation. “Talent” is now abundant and inexpensive.
What’s scarce is knowing what to ask for. Mollick sums it up: “The people who will be successful are those who know what good work looks like and can explain it clearly enough that even an AI can deliver it.”
Organizational Learning as a Competitive Advantage
Those who fail to develop these capabilities will fall behind. The common argument for caution, that technology is evolving too quickly and investments could become obsolete, overlooks a crucial point: Competitive advantage does not come from the technology alone, but from the organizational ability to deploy it. This ability takes time, and that time is running out. Companies that launch pilot projects now are building expertise that will not be available on the market later. They are developing governance frameworks for AI-generated artifacts. They are identifying which use cases work in their specific environment. These organizational insights cannot be copied or bought.
Getting Started: What Sets Successful Pilots Apart
Taking a proactive approach does not mean a full-scale rollout starting tomorrow. It means starting today with a structured pilot program.
1. Choose the pilot team carefully. Start with a team that has a high level of technical expertise on a project with measurable deliverables. Avoid critical legacy systems for the initial phase.
2. Governance from the start. Define clear rules: Which artifacts may be AI-generated—code, documents, analyses, decision templates? How will they be reviewed? Which data flows into which systems? A pragmatic framework with ten guidelines is better than a perfect set of rules that won’t be ready for another six months.
3. Evaluate multiple platforms. The market is evolving rapidly, and the differentiating features of vendors are shifting. A side-by-side comparison of two to three leading solutions over eight to twelve weeks will provide reliable insights for your specific environment.
4. Choose the right metrics. Productivity alone isn’t meaningful. Measure time-to-value, the quality of results, the rework rate, the adoption rate within the team, and which use cases actually work in your environment.
Conclusion: Taking Control Through Early Action
AI agents in software development are no longer a distant dream, and software development is just the beginning. The technology is expanding into marketing, HR, finance, legal, operations, and other areas. A window of opportunity is opening up for IT decision-makers: Those who get on board in a structured way now will actively shape the framework. Those who wait will be forced to react under pressure, with less experience, less time, and fewer options for action.
The recommendation: Launch a pilot project this quarter. Not as a technical experiment, but as a strategic initiative. The organizational lessons learned in the coming months will make all the difference, regardless of which platform ultimately prevails in the long run.
noventum consulting GmbH
Münsterstraße 111
48155 Münster