Agentic AI as a shift

Agentic artificial intelligence is being cast as software engineering’s next big reorganization, with developers moving from writing every step to supervising systems that plan and act. (technologyreview.com) MIT Technology Review Insights wrote on April 14, 2026 that software engineering has already gone through two major shifts this century: open source, then DevOps and agile. The report says a third shift is forming around “agentic AI,” where systems handle more than isolated coding tasks. (technologyreview.com) That framing is spreading through industry posts, including one by Nayeem Ahmad that pointed readers to the broader “redefining software engineering” discussion on X. The idea is not that copilots autocomplete code faster, but that agents can manage longer chains of work across a project. (x.com, technologyreview.com) The same MIT Technology Review Insights piece says 51% of software teams are already using agentic AI in at least limited ways, and 45% plan to adopt it within 12 months. It also says half of organizations now rank it as a top software-engineering investment priority, rising to more than four-fifths in two years. (technologyreview.com) An agent, in plain terms, is a model that does not just answer a prompt but can call tools, fetch data, and take actions. OpenAI’s current API docs describe tools including web search, function calling, remote Model Context Protocol servers, shell access, and computer use for “agentic workflows.” (developers.openai.com) That changes the engineering problem from writing one program to coordinating many moving parts. Model Context Protocol, or MCP, is one example of the new plumbing: its docs describe a client that connects a model to external servers and lists the tools those servers expose. (modelcontextprotocol.io) Vendors are also building more managed versions of that stack. Anthropic’s current platform docs say its “Claude Managed Agents” product lets developers deploy autonomous agents in stateful sessions with persistent event history, instead of hand-building every tool loop. (platform.claude.com) As agents get tool access, governance moves closer to the center of engineering work. Microsoft said on April 2, 2026 that its open-source Agent Governance Toolkit was built for autonomous agents that can book flights, execute trades, write code, and manage infrastructure, and that it maps to the OWASP Top 10 for Agentic Applications for 2026. (opensource.microsoft.com, genai.owasp.org) Microsoft’s post lists risks including goal hijacking, tool misuse, identity abuse, memory poisoning, cascading failures, and rogue agents. Its answer borrows older software ideas — policy engines, identity layers, and circuit-breaker style controls — and applies them to agents at runtime. (opensource.microsoft.com) Standards bodies are catching up more slowly. The National Institute of Standards and Technology said its Artificial Intelligence Risk Management Framework was released in January 2023, added a generative artificial intelligence profile in July 2024, and published a critical-infrastructure concept note on April 7, 2026. (nist.gov) The gap between easier agent building and slower governance is why orchestration, tool interfaces, and oversight are now being discussed as core engineering work. If this framing holds, the next software bottleneck will be less about writing code and more about deciding what autonomous systems are allowed to do. (technologyreview.com, developers.openai.com, opensource.microsoft.com)