Lore

Workflow Entry Points Justify a Native Agent

Linear's team argues a chatbot interface alone gives users no reason to prefer a product's native agent over a generic assistant like Claude or ChatGPT — the interface isn't the differentiator. What justifies building a native, domain-specific agent is decomposing the user's actual end-to-end workflow to find integration "entry points": places where the agent can act inside the tools people already use (e.g., Slack, email, a sales-call transcript) rather than requiring them to go open a chat window. Linear ships multiple such non-chat entry points into existing workflows for this reason.

This is why Linear's leadership isn't worried about agent/MCP usage cannibalizing their polished UI: the product is "multiplayer" — different user types (e.g., support staff working entirely through Zendesk/Intercom) legitimately only ever touch one narrow surface, so being "opinionated" about the product doesn't require funneling everyone through the same interface. See Dynamic Skills Loading (Agent Meta-Tool Pattern) for how the agent scopes what it can do at each of these entry points, and The Agent Line for how Linear expects the range of an agent's responsibility at each entry point to grow over time.

Native Build Enables Tight Skills Control

Linear's agent is built directly into the product rather than shipped as a CLI or an MCP server with skills bolted on externally. That native placement is what lets the team dynamically load hundreds of Dynamic Skills Loading (Agent Meta-Tool Pattern) skills with tight, opinionated control over behavior — a level of control that's hard to retain when skills are exposed through a generic external protocol. Native architecture and deep skills control reinforce each other: the more entry points and skills a product wants to support, the stronger the case for building the agent in rather than bolting one on.

On-Ramps and the Subsystems Framing

On-ramps: decomposing the workflow

Because "no one sits down and one-shots their whole job," the practical design method is to map a user's actual end-to-end workflow and find the natural points where an agent can plug in — Slack threads, meeting debriefs, project docs — rather than funneling everything through one chatbot. Each of these entry points is an "on-ramp" into the agent.

Collection of subsystems, not one chatbot

The resulting product is described internally as a collection of subsystems/integration points rather than a single chatbot. The in-app chat interface is necessary but explicitly secondary — most value comes from the distributed on-ramps embedded where work already happens; chat mainly handles follow-up. This also explains why a native agent beats exposing the same actions purely via Model Context Protocol (MCP) to third-party agents: MCP lets any external agent call your tools, but it can't embed your own product's on-ramps or opinions into users' existing workflows the way a native, workflow-aware agent can. See also third-party-plugin-escalation-pattern for how peripheral users get their own on-ramp.

"AI is not an end-to-end solution, it's a middle's solution" — the agent slots into the middle of an existing workflow rather than replacing the whole workflow.