Building the platform for agent coordination — Tom Moor, Linear

Introduction and Linear's AI Journey 00:01

  • Tom Moor introduces himself as the engineering lead at Linear, presenting Linear’s story with AI, the features built, and views on the future of software development.
  • Linear started as an issue tracker, evolving into an operating system for engineering and product teams, focusing on speed, clarity, and friction reduction.
  • By early 2023, Linear initiated a small internal team to experiment with AI, particularly for summarization and similarity features, without prior AI experience.
  • Early efforts centered around search, leading the team to experiment with vector databases, ultimately settling on OpenAI embeddings stored in PG Vector on GCP for pragmatic reasons.
  • Initial AI-powered features included “similar issues” suggestions using basic cosine embedding comparisons, natural language filters for issue navigation, and automatic issue creation from Slack threads.
  • A co-pilot feature was attempted but not shipped due to not meeting Linear’s quality bar.

Advancements in AI Capabilities and Platform Foundations 04:44

  • By late 2024, increased AI capabilities (larger context windows, planning and reasoning models, multimodal APIs) led Linear to rebuild its search index.
  • Transitioned to hybrid search using Turppuffer, moved embeddings from OpenAI to Coher for improved performance, and completed a large-scale backfill of embedding data.
  • These advancements established a more robust search foundation for new features and integrations.

Next-Generation Features and Intelligence 06:13

  • Developed “product intelligence,” an improved similar issues feature using a pipeline with query rewriting, hybrid search, reranking, and deterministic rules.
  • This feature maps relationships between issues, explaining how and why they relate, enabling suggested labels, assignees, possible duplicates, and relevant project/team matches.
  • Enables efficient triage for large organizations processing many tickets.

Customer Feedback Analysis and Workspace Updates 07:24

  • Added customer feedback analysis, leveraging LLMs to synthesize feedback from multiple channels and recommend features or projects, reportedly outperforming 90% of job candidates in analysis tasks.
  • Introduced daily or weekly "pulse" features that summarize workspace updates, available as audio via the mobile or desktop app for convenient consumption.

Automation from Media and Bug Reporting 09:09

  • Created an “issue from video” feature that parses video bug reports to identify reproduction steps and automatically generate issues, saving time and effort.

The Agent Platform: Coordination and Extensibility 09:39

  • Linear aims to make the platform pluggable to accommodate diverse team workflows, introducing the concept of agents as scalable, cloud-based teammates.
  • Recently launched a platform for agent integration, supporting agent orchestration alongside human users within the same workspace.
  • Demonstrated integration with coding agents (e.g., Codegen, Charlie) capable of planning, creating PRs, conducting root cause analysis, and linking issues to code.
  • Bucket, a feature flagging platform, is integrated as an agent that can create and manage feature flags directly in Linear.
  • Discussion of PM agents and Intercom's Finn agent, which can automate customer replies when issues are resolved.
  • Working on richer agent surfaces: making agent deliberations and tool calls visible, allowing users to interrupt agents mid-task, and supporting agents across different roles.

Impact of Agent-Driven Automation 13:54

  • Widespread agent adoption will reduce backlog size and eliminate excuses for unresolved issues, with agents able to process and potentially resolve a significant portion automatically.
  • Anticipates increased productivity and quality as agents take on repetitive or menial tasks, enabling teams to build more and faster.

Agent Architecture and Integration Details 14:39

  • Agents in Linear have full user identities, histories, and audit trails; installed via OAuth, managed by admins, and operate transparently.
  • Linear’s mature GraphQL API and granular scopes allow agents to perform any task a human can, with new webhooks and mention/assignment controls.
  • Upcoming SDK will simplify agent integration, augmenting the existing API and enabling easier development.

Best Practices for Building and Integrating Agents 16:19

  • Agents should respond quickly and precisely, acknowledging requests with clear, context-aware replies.
  • Developers should ensure agents feel native to the platform they’re operating in, adapting to platform language and conventions.
  • Encourage natural behaviors: agents should update issue states, follow up on messages, and act as reliable team members.
  • Agents should clarify intent before acting, often by proposing a plan and seeking feedback first.
  • Emphasize being concise, relevant, and value-adding—avoiding excessive, unnecessary language in interactions.
  • Aim for agent behavior that mirrors effective human teammates, prioritizing utility and clarity.

Conclusion and Collaboration Invitation 19:24

  • Tom closes the talk by inviting interest in building or integrating with Linear’s agent platform.