New engineers orchestrate AI agents

AI agents are changing software work from line-by-line typing to assigning tasks, feeding context, and checking what comes back. (blog.jetbrains.com) (docs.github.com) That shift is now showing up in vendor docs and surveys, not just social posts. JetBrains said in its 2025 Developer Ecosystem Survey that 85% of developers regularly use AI tools, and 62% use at least one coding assistant, agent, or AI code editor. (blog.jetbrains.com) The same survey said developers mostly want AI to handle boilerplate code, internet search, code conversion, documentation, and change summaries. JetBrains said developers still prefer to keep debugging and application-logic design under human control. (blog.jetbrains.com) In plain terms, an AI coding agent is less like autocomplete and more like a junior teammate that can read files, run tools, and propose patches. Orchestration means the engineer decides the goal, the context, the sequence of steps, and the checks before anything ships. (claude.com) (langchain.com) LangChain described that next layer as a “control plane” for software delivery, where multiple agents share memory, state, and traceability across requirements, testing, security, and deployment. In a Cisco pilot described in the post, coordinated agents cut time-to-root-cause by 93% across more than 20 debugging workflows and saved more than 200 engineering hours across 512 sessions in one month. (langchain.com) Anthropic’s recent product posts point the same way. In April 2026 alone, the company published updates on managed agents, parallel agents on desktop, routines, built-in memory, and coordination patterns for multi-agent work. (claude.com) The IDE has not disappeared in this model; it becomes the place where humans inspect context and make final edits. A JetBrains post by LangChain engineer Jacob Lee said he now makes “point edits and suggestions” instead of writing modules by hand, but still uses the IDE to understand scope, review branches, and point agents at the right files. (blog.jetbrains.com) GitHub’s guidance is even more explicit that human review remains mandatory. Its Copilot documentation says teams should run tests and static analysis first, check whether generated code fits project architecture and conventions, and reject code that is harder to refactor than to rewrite. (docs.github.com) Microsoft has been building that review layer into its own workflow. The company said in July 2025 that an internal AI review assistant was already supporting more than 90% of pull requests across Microsoft, affecting more than 600,000 pull requests per month, while human reviewers focused on architectural and security questions. (devblogs.microsoft.com) The caution is that more generated code does not automatically mean better software. JetBrains listed inconsistent code quality, limited understanding of complex logic, privacy and security risks, and lack of context awareness among developers’ top concerns. (blog.jetbrains.com) Google’s DORA research put the same warning in organizational terms. Its 2025 report said AI acts as an amplifier, strengthening teams with good systems and magnifying dysfunction in teams that already struggle with process, review, and measurement. (dora.dev) (research.google) The result is a different definition of engineering productivity. The scarce skill is moving from typing every function yourself to framing the problem, supplying the right context, choosing the right agents, and catching the failures before users do. (docs.github.com) (langchain.com)