Skip to content
Product Agents

Search Product Agents

Start with a guide, or search a topic

    New here? Read the five guides in order

    Subscribe

    Linear Agent

    Linear charges for people and passes agent work through at cost.

    Linear’s agent takes delegated issues, writes code and runs unattended Loops, while the person assigned stays responsible. Here is how it works, what it costs, and what “complete” does and doesn’t mean.

    The Linear Agent composer on a dark background, with an “Ask Linear anything…” field and suggested prompts: What changed this week?, What needs my attention today?, and What’s at risk right now?
    Image: Linear documentation

    At a glance#

    What it does

    Answers questions about the workspace, creates and updates issues and projects, and drafts project updates. In a coding session it writes code and opens a pull request from an issue. Loops run the same agent on a schedule or when something changes.

    Price

    Chat is included on every plan, Free ($0) upward; seats are $10 (Basic) or $16 (Business) per user per month, billed yearly. Coding sessions and Loops draw prepaid AI credits. Those bill model tokens at the provider’s published rates with no markup, plus $0.25 per 20-minute block of sandbox time. Loops are on Business and Enterprise. Credits expire after 12 months.

    Model

    Coding sessions run Claude Code or Codex. The default model under Auto is GPT-6 Sol. Claude Opus 5.5, Claude Fable 5, Claude Sonnet 5 and the open-weight GLM 5.3 are among the options. Linear says chat is “powered by frontier language models” and doesn’t name them.

    Takes actions?

    Yes. In chat it acts within the asking person’s permissions. On an issue it works as a delegate while a person stays the assignee. A coding session ends in a draft pull request that a person merges. Loops write on their own, with permissions set per loop.

    Claimed results

    Linear says agents are installed in 95% of paid workspaces and create 50% of the work, up from 3% a year earlier. Internally, its triage-and-fix flow resolves about 30% of incoming bug reports. Linear cites Ramp, where agents write more than 60% of pull requests.

    Launched

    Third-party agents as teammates (Devin, ChatPRD, Codegen) May 20, 2025. Linear Agent public beta March 24, 2026. Coding sessions June 11, 2026. Loops July 20, 2026. Pass-through pricing for coding sessions August 20, 2026.

    Owner / Status

    Linear, private. It passed $100 million in annual recurring revenue in 2026 and has more than 40,000 paying companies. In August 2026 it ran a $99 million employee tender at a $2.5 billion valuation, with net revenue retention at 177%. CEO Karri Saarinen told Devclass in March that “issue tracking is dead.”

    Why it matters#

    Linear is the issue tracker more than 40,000 companies pay for, and its agent is built into that record. An issue keeps its human assignee while an agent works it as a delegate. Chat with the agent comes with every seat, even on the Free plan. Agents aren’t seats at all.

    Linear charges for people and passes the agent’s coding work through at cost. Coding sessions bill model tokens at the provider’s published rates, with no markup, plus $0.25 per 20 minutes of sandbox time. A person stays the assignee on every issue an agent works, and Cursor, Codex and GitHub Copilot can work beside Linear’s own agent.

    How it works#

    What happens when a request reaches Linear Agent.

    1. Report

      A bug lands in the team’s triage queue. A teammate delegates the issue to Linear, or a Loop set to watch triage does it. The teammate stays the assignee, Linear is the delegate, and the handoff is written to the issue’s Activity feed.

    2. Investigate

      Linear Agent reads the issue, its history and the discussion, then looks at the codebase. Through MCP it can pull evidence from tools like Sentry or Datadog.

    3. Write

      The agent starts a coding session in a managed sandbox, using Claude Code or Codex. It installs what the repository needs, runs the app, and tests its change in a browser before handing it back.

    4. Return

      Linear drafts a pull request, adds the diff and screenshots to the issue’s Reviews tab, and tags an engineer. The session shows one of six states the whole time: pending, active, awaiting input, complete, error or stale.

    5. Review and merge

      The engineer reads the diff, asks for changes in a comment, and merges from Linear. “Complete” only means the session ended. The issue stays with its assignee until a person closes it.

    6. Pay

      The workspace’s AI credits drop by the tokens used at the provider’s rate plus $0.25 per 20-minute sandbox block. An admin tops up from $10, or sets auto top-up from $50. At zero, no new session starts.

    The decisions worth studying#

    What they chose, why, and what it costs.

    1The person stays the assignee. The agent is a delegate.

    An issue in Linear is assigned to one person at a time. When that person hands the issue to an agent, Linear adds the agent as a delegate and leaves the assignee in place. Linear’s docs put it plainly: the assignee “remains responsible for the work, while the agent contributes on their behalf.” Delegated issues still show in the assignee’s My Issues list. Every handoff is written to the Activity feed with who made it.

    Linear backs the split with identity. An installed agent signs in as an app of its own and doesn’t borrow the installer’s account. App agents can’t ask for admin rights, sign in to the app, or manage users. They don’t count as billable seats.

    The cost is in the status words. A session shows pending, active, awaiting input, complete, error or stale. Complete means the agent stopped. Nobody has accepted anything until the assignee says so.

    Linear · agent session on an issueProduct screenshot
    An agent session on a Linear issue: Working for 04:35 after a teammate’s mention, with a link to the agent’s own interface.
    A delegated session on an issue. The state sits on the card, and the agent’s own interface is one link away. Linear

    2Agent compute is sold at cost, and the seat is the product

    Linear Agent chat is included in the seat price on every plan. The work the agent does on its own is metered: coding sessions and Loops draw from prepaid AI credits. On August 20, 2026, Linear set the rate. Model tokens are billed at the provider’s published rates “with no markup,” plus $0.25 per 20-minute block of sandbox time. A typical Loop run costs $0.07 to $0.20. Admins top up from $10, and funds expire after 12 months.

    Linear’s stated reasoning, from the March launch: chat stays in the seat, and high-volume compute may move to usage pricing. My read: Linear’s margin lives in seats and expansion, which it reports as 177% net revenue retention. Agent compute is a cost it passes through, the way a cloud bill is. That lets it welcome rival agents and never tax a customer for using them.

    The risk sits with the buyer. Credits expire, and at a zero balance nothing new starts.

    3Its own coding agent sits beside Cursor, Codex and Copilot

    Linear opened the door to other agents first. On May 20, 2025 it let Devin, ChatPRD and Codegen join workspaces as teammates. Cursor followed on August 21, GitHub Copilot on October 28 and OpenAI Codex on December 4. By March 2026, the CEO told Devclass, coding agents were in 75% of Linear’s enterprise workspaces. Linear shipped its own coding sessions on June 11, 2026, running Claude Code or Codex in a sandbox it manages.

    Linear is open about where its agent stands. The default model under Auto is GPT-6 Sol; Claude Opus 5.5, Claude Fable 5 and the open-weight GLM 5.3 are a setting away. Since September 24, Auto routes simple tasks to a faster model. Linear sells the sandbox and the record; the model is a choice.

    The cost: a third-party agent reports what its maker chooses. Linear sets the contract: acknowledge within 10 seconds, emit thought, action, elicitation, response or error, and halt on stop. How much detail arrives is up to each provider.

    4The draft lands where the finished work lives

    Ask Linear Agent for a project update and the answer doesn’t stay in chat. In Linear’s June 18, 2026 recording, the author types “Help me write an update for this project.” The agent plans to gather what changed since May 1 from issues, project activity and the connected Slack channel. The draft then appears in Linear’s own update composer, marked On track, with Cancel and Post update underneath.

    The composer shows what the prose hides. The text says the onboarding experience is “fully rolled out.” Right below, the GA milestone reads 3% to 33%. The author decides which claim posts. When it posts, the update carries the author’s name and date, and the header doesn’t mention the agent.

    My read: this is the right place for an agent’s draft, and it moves the burden. The person’s name is on the claim, so the review has to be real.

    Linear · project update composerProduct screenshot
    Linear’s update composer with Comment and Update tabs and On track, two paragraphs beginning “The refreshed onboarding experience is now fully rolled out”, then Status In Progress → Maintenance, Target date Apr 2026 → May 2026, Progress since May 1 with GA 3% → 33%, and Write with Agent, Cancel and Post update.
    The agent’s draft in Linear’s own composer. The text says “fully rolled out”; the milestone says 33%. Linear changelog recording · June 18, 2026 · 0:12

    What to watch#

    • “Complete” says nothing about acceptance

      A session marked complete has ended. Keep a separate accept step for the assignee, and read stale as “Linear stopped hearing from the agent,” which is different from the agent stopping.

    • Credits are prepaid and they expire

      AI credits expire 12 months after purchase, can’t be refunded or moved between workspaces, and at zero no new coding session or Loop starts. The promotional Loops credits end December 31, 2026.

    • Loops write unattended

      A Loop can be allowed to change entities beyond the one that triggered it, query any website, and start coding sessions. Nobody confirms a scheduled run; the run history is the only review. Restoring a Loop version restores its settings. What a past run changed stays changed.

    • Stop halts the agent; undo isn’t documented

      On a stop signal the agent must make no further changes and report that it stopped. Reversing what it already changed in a repository or another tool is left to you.

    What it means for your team#

    What to decide before you ship an agent like Linear Agent.

    • Product

      Keep who owns the work and who does it as two fields on the same item, and add an accept step for the owner. A session marked complete has only ended.

    • Finance

      Agent compute is a prepaid, expiring balance, charged at provider token rates plus $0.25 per 20 minutes of sandbox. Decide who can top up, set an auto top-up floor, and watch the usage dashboard.

    • Security

      Give each agent its own identity, as Linear does: installed agents sign in as an app, can’t request admin scope, and can’t manage users. Review each Loop’s permissions, since a Loop can write beyond the entity that triggered it.

    • Platform

      If you build an agent for Linear, answer the webhook within 5 seconds, emit a thought within 10, and halt on stop with a final response. Test what a stopped session leaves behind, because undo isn’t documented.

    Steal this

    1. Keep who owns the work and who does it as two fields on the same item.

    2. Put attended help in the seat price and meter unattended runs at a published rate.

    3. Show the owner a status that says whether to act, in place of a spinner.

    Reference#

    The evidence behind this teardown: recordings, sessions, what Linear Agent gets right and wrong, and every source.

    Walkthrough · 3 frames

    From “Help me write an update” to a posted project update#

    Frames from the recording in Linear’s changelog entry “Agent assisted project updates”, June 18, 2026. A vendor recording, not a live test.

    1. Step 1 of 3: The author asks for a project update

      Linear’s agent panel titled “Write Onboarding 2026 project update”: the request “Help me write an update for this project”, “Onboarding 2026 added to context”, “Thinking...”, and the plan “I’ll gather what changed since the May 1 update across completed issues, project activity, and the connected Slack channel, then summarize the facts before drafting.”
      Image: Linear changelog recording · June 18, 2026 · 0:05

      1. The author asks: “Help me write an update for this project”.

      2. Linear adds the project, Onboarding 2026, to the agent’s context.

      3. The agent plans to gather changes since the May 1 update from issues, project activity and Slack; the recording doesn’t show the retrieval.

    2. Step 2 of 3: The draft lands in Linear’s update composer

      Linear’s update composer with Comment and Update tabs and On track, two paragraphs beginning “The refreshed onboarding experience is now fully rolled out”, then Status In Progress → Maintenance, Target date Apr 2026 → May 2026, Progress since May 1 with GA 3% → 33%, and Write with Agent, Cancel and Post update.
      Image: Linear changelog recording · June 18, 2026 · 0:12

      1. The draft appears in Linear’s own update composer, on the Update tab, marked On track.

      2. The prose says “fully rolled out” while GA progress reads 3% → 33%, two claims for the author to reconcile before posting.

      3. The composer offers Cancel and Post update: a chance to review before the update goes out.

    3. Step 3 of 3: The update posts under the author’s name

      The posted project update: On track, meg, Jun 16, and the same two paragraphs beginning “The refreshed onboarding experience is now fully rolled out”.
      Image: Linear changelog recording · June 18, 2026 · 0:15

      1. The update posts under the author’s name, dated Jun 16; the header doesn’t mention the agent.

      2. The posted text matches the draft word for word; the recording doesn’t check the facts against the project’s issues.

    Sources#

    1. Linear pricinglinear.app
    2. AI creditslinear.app
    13. Linear Agentlinear.app
    14. Coding sessionslinear.app
    15. Loopslinear.app
    20. Signalslinear.app

    Sources checked September 28, 2026.

    Next teardownNotion Agent

    Notion has two kinds of agent: a personal Agent that works with the person’s own access, and Custom Agents that have their own permissions and can run on schedules or when something changes.