30 秒でわかる
As AI agents write more code, engineers still have to decide what to build and verify that it works, says Microsoft executive Charles Lamanna. He says coordination—not typing code—already consumes much of the time needed to ship a feature. He recommends setting a broad direction, testing competing ideas and trusting teams to execute. He also argues that workplace AI should support people and agents working together, rather than copy the solitary coding workflow into every job.
- 1Agents change the demands on software interfacesLamanna expects agents acting for people to become major software users. He says they can handle much larger search results than people and will put pressure on performance, cost and reliability.
- 2Copilot and Office need to work in both directionsMicrosoft embeds Office editors beside the Copilot conversation and puts Copilot in Office apps. Lamanna says a mouse remains convenient for small, precise changes.
- 3Solitary coding is a poor model for all office workSales and operations alternate between independent work, communication and collaboration. Lamanna describes shared workspaces where colleagues and agents contribute to the same goal.
- 4Meeting records can become context for later agent tasksAlthough people rarely replay every meeting, Lamanna says transcripts can help an agent recover the reasons behind a decision. He cautions against generating long documents for colleagues without reviewing them.
- 5Engineers who write less code still have to think hardLamanna says engineers using coding agents must design systems, verify and test outputs, manage safe releases and monitor results. He compares the change with earlier shifts to compilers and cloud services.
- 6As coding speeds up, coordination becomes the constraintLamanna says leaders must align people on strategy and quality. His approach combines a written high-level direction, experiments to compare options, and giving trusted teams room to execute.
ポイント
- Set goals and guardrails, verify generated code, and monitor its rollout; producing code is only part of engineering.
- Lamanna says coordination around priorities and quality is a major constraint even when implementation becomes faster.
- Test competing product ideas with experiments instead of settling every disagreement in a meeting.
- Design workplace AI for shared projects and human collaboration, not only an individual managing agents.
- Keep the familiar editing interface for precise adjustments while using AI for broader changes.
チャプターごとに整えた全文を読む
Agents become demanding software users
Lamanna expects agents directed by people to account for a large share of software use. Their demands for speed, reliability and machine-readable results can change interfaces.
Charles Lamanna argues that workplace AI has been too closely modeled on software engineers working alone. Sales and other jobs depend on frequent human interaction. He also says engineers using agents are still solving problems, even when they write less code themselves.
Peter Yang asks whether agents will become software’s primary users. Lamanna thinks they will be major users of many systems: in his own work, agents generate code, test apps and debug, while he directs the tasks, reviews the results and verifies them.
As a leader at Microsoft, Lamanna also uses Copilot’s long-running agents for tasks such as reviewing GPU capacity, planning budgets and thinking through the next quarter’s roadmap. They can connect to software through interfaces or use a browser and click through an application.
He expects many agents to act on people’s behalf. As an illustration of the load this can create, he cites a GitHub report of a fourteenfold increase in load over a year, which he attributes to agents participating in work rather than a comparable increase in human developers.
Agents are demanding customers, Lamanna says: they need efficiency, low cost, performance and reliability. An interface designed for a person may be a poor fit for a machine carrying out many small operations.
A search page for a person might return ten links with short snippets. An agent can process a thousand results with longer excerpts, so software interfaces may develop different modes for agent clients.
Why Office and Copilot work side by side
Microsoft integrates Office editors into Copilot and Copilot into Office. Lamanna expects prompts to complement familiar interfaces rather than quickly replace them.
Yang asks how Microsoft is moving applications designed for people clicking through menus toward an AI-led experience. Lamanna describes three strands of work, starting with embedding full Office editors alongside the Copilot conversation.
A person can ask for broad changes in the prompt box and then use the familiar editor for small ones. Moving a text box eight pixels or selecting text for formatting can be faster with a mouse than a written instruction, he says.
Microsoft has also brought Copilot into Word, Excel and PowerPoint. Someone spending hours on a presentation may prefer to stay in PowerPoint and switch between direct editing and prompts as needed.
The Office family has kept expanding as technology changes, Lamanna says. He points to Pages, an HTML-oriented format with Office sharing and security, and the Today and Workspace experiences built by the Office team.
Lamanna expects more people to begin creating documents with a prompt, but does not think that makes the applications unimportant. Yang agrees that people may still want to make the last small adjustments to a slide themselves.
Design AI around shared work
Lamanna says many jobs rely on frequent collaboration rather than isolated work. Microsoft’s shared workspaces are intended to combine people, agents and documents around a goal.
Yang asks about multiplayer AI: managing agents alone can be isolating. Lamanna says coding often involves an individual branch followed by review, but sales, customer service and operations move much more frequently between individual work, conversation and coordination.
Microsoft wants Copilot to support multiple people and agents together. In the Office experience, a person can work on a document with Copilot while colleagues have the same file open and coauthor it in real time.
Workspace groups people and agents around a project, such as preparing a customer proposal or responding to a supply-chain disruption. Lamanna says it can suggest people and agents to involve while users switch between their own tasks and shared work.
Keep useful context without burdening coworkers
Meeting transcripts and older files can help agents reconstruct decisions. Lamanna says workers should review AI-generated documents and avoid sending unnecessary material.
Yang asks how documents should serve both people, who need a concise answer, and agents, which might read a lengthy appendix. He also asks how to prevent people from flooding colleagues with AI-generated documents.
Lamanna separates reusable data from material sent to people. Meeting recordings and transcripts are rarely worth replaying in full, but an agent can use them to find why a decision was made or which alternatives were considered.
Old email and files can likewise become useful context even if nobody opens most of them. Lamanna says Microsoft’s Work IQ and Copilot can retrieve material at the occasional moment it is needed.
For communication between people, he argues for restraint. If someone would not take the trouble to write ten pages, they should not generate ten pages and send them around without care. He expects workplaces to develop norms for reviewing and disclosing AI assistance.
Engineering shifts toward design and verification
Coding agents change how engineers work rather than removing the need to think. Lamanna compares the transition with earlier advances in programming languages and cloud computing.
Yang asks whether people’s job is becoming merely training agents, noting that some engineers miss solving problems by writing code. Lamanna says the work changes but remains more than onboarding a machine.
Engineers using automated coding systems still make architectural choices, decide what signals to give agents, verify generated code, test it and plan safe releases. Some tell him they have never had to think harder, even though they are exercising a different skill.
Lamanna compares this with earlier changes: compilers reduced hand-written assembly, garbage collectors removed much manual memory management, and cloud services reduced server maintenance. Each moved some work to a higher level of abstraction.
He expects people to set goals, ambition and boundaries, verify the work and keep monitoring it. For his own decisions, AI can now prepare a dashboard from data that once took him considerable time to slice and plot.
Coordinate teams and test competing ideas
Lamanna sees alignment on priorities and quality as a major part of shipping. He writes a broad strategy, uses experiments to settle questions and gives teams autonomy.
Yang recalls lengthy product requirements documents at Microsoft and asks how a large organization stays aligned when building becomes much faster. Lamanna answers that writing code and shipping a feature take less time than agreeing on priorities, strategy and the standard of quality.
Every six months, Lamanna writes roughly twenty pages laying out his high-level thinking. He gives the example of supporting multiple models and routing requests to provide quality at a lower price; teams can use that direction without being told to build each individual feature.
His second approach is experimentation. Rather than argue indefinitely about a design, a team can run several variants and compare user satisfaction and behavior, as search teams have long done with results layouts.
Third, leaders need to hire people they trust and give them authority to act. Lamanna says that remains important as AI takes on more engineering work.
Lamanna says the current pace of change is the fastest he has seen in a career that began around the cloud boom. He points to the challenge of adapting Office and Microsoft 365 while maintaining software on which organizations depend.
同じチャンネルの記事
17:57EN
22:12EN
42:01EN
25:17EN
21:47EN
46:40EN