Linear · AI product · Mid · 45 to 60 minutes

How should Linear constrain its AI agent?

Linear is a software product for planning, tracking, and collaborating on issues, projects, and related work. Answer this product management case and get it graded line by line against a weighted rubric.

AI product judgment
Scope definition
Risk management
Systems thinking
Written communication

The situation

Linear is preparing an early beta of an AI agent that can help users do work inside Linear and from Slack. The CEO and head of engineering ask you, the PM for the effort, to recommend how much freedom the agent should have before the team hardens the architecture and starts adding more capabilities. The promise is clear: users want the agent to handle messy, underspecified requests like drafting updates, creating or linking work, and pulling signal from customer feedback. The risk is also clear: if you script every path, the agent becomes a thin macro layer, but if you expose too much of Linear's underlying data model and tools, it may take surprising or unsafe actions. The team also knows that many requests are short, context varies by surface, and the product already has many distinct concepts and operations. You need to propose a product strategy for boundaries, capability exposure, approvals, and execution flow that can support an initial beta and expand over time without making every interaction slow, costly, or brittle.

What to produce

Write a 1 to 2 page product strategy memo.

Your answer must

State 3 product principles that separate where the agent should behave consistently from where it should stay flexible.
Recommend how the agent should receive capabilities and product context for a task, including how this should scale as Linear supports more features.
Propose a policy for risky actions and user confirmation, with at least 3 concrete examples of when approval is and is not required.
Compare at least 2 approaches for the action layer and execution stack, such as broad low-level access versus curated capabilities, and choose one with trade-offs.
Define beta success using 3 metrics or review signals, including at least 1 for usefulness and 1 for safety.

Constraints

The agent must work in both Linear and Slack.
User prompts are often short and ambiguous.
Some actions are hard to undo or may affect public synced threads.
Loading all tools and instructions into every run is not viable.

Format

Use short sections with bullets, and include a simple trade-off table if helpful.

How Linear actually played it

Linear explains how it designed its Agent, noting that AI inverts the usual goal of making software behave consistently: the value lies in the agent doing work that was not explicitly scripted. The post covers how much to constrain behavior. A thoughtful read on product design with nondeterministic features.

Overscripting an agent removes its value.
Decide where consistency matters and where flexibility does.
AI features need different design principles than rules-based ones.

Built from How we built Linear Agent by Matthijs Wolting.