How to use AI agents as a product manager
Seven jobs a product manager can hand to AI agents today, the tools for each, what to check before trusting the output, and the one job that stays yours.
A product manager asks an engineer how the free trial actually ends. The engineer is busy, so the answer comes two days later and is half right. Meanwhile an AI coding agent on the same team could have read the code and answered in a minute.
AI agents are not only for engineers. Anthropic reports that teams across the company use Claude Code, including product design, growth marketing, and legal, and describes agentic coding as dissolving the boundary between technical and non-technical work. This guide covers seven jobs a product manager can hand to agents today, the tools for each, and what to check before trusting the result. Sources were checked on 15 November 2026.
Seven jobs to hand to an agent
| Job | Tool | What you get | What to check |
|---|---|---|---|
| Understand how a feature works | Claude Code in plan mode | A plain-English explanation from the code | Ask for the file and function behind each claim |
| Map edge cases before a build | Claude Code, Figma MCP | A list of error states, empty states, and flows | Whether each case matters to users |
| Prototype from a design | Claude Code with Figma MCP | A working prototype to click through | That it is a prototype, not shippable code |
| Query product data | PostHog MCP | Answers from your own analytics | The query it ran, not only the number |
| Draft an agent-ready spec | Any agent, a PRD template | A first draft with numbered requirements | Every requirement, edge case, and non-goal |
| Triage tickets | Linear or Atlassian MCP | Grouped, summarized, deduplicated tickets | Anything it closed or changed |
| Automate repetitive work | Claude Code with subagents | A script or workflow that runs on its own | A sample of the output, every time |
1. Understand how a feature works
The fastest win is asking the code directly. Open Claude Code in the repository and ask: "How does a free trial end? What happens to the user's data?" Use plan mode, so Claude explains and proposes without editing anything.
Anthropic's article describes new hires using Claude Code to get up to speed on codebases, with Claude reading the CLAUDE.md files and explaining how pieces connect. A PM gets the same benefit: answers about current behavior without waiting for an engineer.
Check: ask the agent to name the file and function behind each claim, and spot-check one or two. Code shows what the system does, not what it was meant to do.
2. Map edge cases before the build
Anthropic's product design team found an unexpected use: mapping out error states, logic flows, and system statuses to find edge cases during design instead of during development. The article says this improved their initial design quality.
Give the agent the feature and the relevant code, then ask: "List every error state, empty state, permission case, and limit this feature can hit." You will get cases nobody wrote down. Some matter. Some don't. Deciding which is your job.
3. Prototype from a design
The same team fed Figma designs to Claude Code and let it build and iterate on features. With Figma's MCP server, the agent reads the design as structured data, not a screenshot. For the setup, see how to give AI coding agents design context.
A clickable prototype answers questions a document can't: does this flow feel right, is this step necessary, what did we forget? Work in a separate branch or a copy, and treat the result as a prototype. Shippable code still goes through engineering.
4. Query your own product data
Most PMs wait in a queue for data questions. An agent connected to your analytics can answer many of them directly. PostHog's MCP server lets an agent run HogQL queries, look at feature flags and experiments, and check error tracking, with OAuth sign-in.
Ask: "What share of new workspaces created a second project in their first week, by plan?"
Check: ask the agent to show the query it ran. A wrong filter gives a confident wrong number. If the answer will drive a decision, have someone who knows the data look at the query.
5. Draft an agent-ready spec
If engineers build with AI agents, your spec will be read literally by one. An agent can turn your notes into a first draft with numbered requirements, acceptance scenarios, edge cases, and open questions marked. Then you edit it.
Use a fixed shape so drafts are consistent: the PRD template for AI coding agents follows the structure GitHub's Spec Kit uses. For the workflow around it, see spec-driven development for product managers.
Check: every requirement, non-goal, and open question. The draft is where the agent's assumptions hide.
6. Triage tickets
Connected to your tracker, an agent can group a hundred incoming tickets by theme, find duplicates, and summarize what customers are asking for. Linear's MCP server includes a read-only endpoint, which is the right place to start: the agent can read and summarize without changing anything. Atlassian's server covers Jira and Confluence.
Check: if you give it write access, review every ticket it closes, merges, or reprioritizes.
7. Automate repetitive work
Anthropic's growth marketing team built a workflow that processes CSV files of hundreds of ads, finds underperformers, and generates new variations within character limits, using two specialized subagents. The legal team built a prototype phone tree to route people to the right lawyer.
Look for work you do every week with the same steps: a weekly metrics summary, release notes from merged pull requests, a digest of customer feedback. Those are good first automations. Review a sample of the output every time it runs.
How to start safely
- Start read-only. Plan mode in Claude Code, read-only tracker access, and analytics through your own login.
- Use a branch or a copy for anything that writes code. Never push to the main branch.
- Connect the fewest tools that cover the job. Each connected server is another system the agent can read or change.
- Ask for sources. File and function for code answers, the query for data answers, the ticket IDs for summaries.
- Keep a human in every approval.
What stays with you
An agent can produce options, edge cases, data, drafts, and prototypes faster than any PM. It should not decide which option wins. Choosing what to build, what to cut, and which trade-off to accept is the product manager's job. So is talking to customers and approving what ships.
That job is getting harder to do well, because more of the build is now done by agents. An engineer's agent builds from whatever context it is given. A decision you made in a meeting, or a scope change agreed in a Slack thread, doesn't reach it unless someone writes it down where the agent will read it. When that doesn't happen, the agent builds the old version with full confidence.
Where Lockstep fits
Lockstep makes sure your decisions reach the agents doing the work. Paste a PRD or brief, review the requirements it extracts with their source passages, and approve the ones that are right. Decisions made later in Slack threads or pull requests are captured as drafts for their owner to approve. Once a developer connects their repo, their Claude Code sessions are briefed on the approved requirements and decisions that apply, and an opt-in check can flag code changes that may conflict with one.
You can start without a repo or a developer. Session briefings work with Claude Code today. See Lockstep for product teams.
Sources checked 15 November 2026
- https://www.anthropic.com/news/how-anthropic-teams-use-claude-code
- https://posthog.com/docs/model-context-protocol
- https://linear.app/docs/mcp
- https://developers.figma.com/docs/figma-mcp-server/
- https://github.com/lockstep-team-agent/lockstep
Related reading
For writing specs agents can follow, see spec-driven development for product managers. For a template, see the PRD template for AI coding agents. For checking what was built, see how do I know engineering built what the PRD said?. For deciding well in the first place, see product decision-making frameworks. For a general PRD layout, see the product requirements document template. For where prototyping ends and engineering begins, see vibe coding vs agentic coding. For which Claude agent fits which job, see Claude Code vs Cowork. For using AI in planning meetings, see sprint planning with AI. For the difference between agents and other AI, see agentic AI vs AI agents vs generative AI.
Frequently asked questions
- How can a product manager use AI agents?
- Hand them work with a clear output you can check: explaining how a feature works in the code, listing edge cases and error states before a build, prototyping from a Figma design, querying analytics, drafting a spec, or triaging tickets. Keep prioritization, trade-offs, and approvals with people.
- Do product managers need to code to use Claude Code?
- No. Anthropic reports that teams outside engineering at Anthropic use Claude Code, including product design and growth marketing. Start in plan mode or a read-only setup so the agent can read and explain without changing anything, and work in a copy or branch when you prototype.
- What should a product manager not delegate to an AI agent?
- Decisions. An agent can lay out options, edge cases, and data, but choosing what to build, what to cut, and which trade-off to accept is the PM's job. So is talking to customers and approving what ships.
- Which MCP servers are useful for product managers?
- The ones connected to where product work lives: your tracker (Linear or Atlassian), your analytics (PostHog), your design files (Figma), and Slack. Most are hosted by the vendor and use OAuth, so the agent gets your access and no more. Linear also offers a read-only endpoint.
- How do AI agents change the product manager's job?
- More of the build is done by agents that read whatever context they are given. A decision that lives only in a meeting or a Slack thread never reaches them. The PM's growing job is making sure product decisions are written down, current, and in front of every agent before it builds.
Keep reading