August 2026 Summaries

Linear completed a $99 million tender offer at a $2.5 billion valuation, allowing current and former employees to sell some vested equity without the company raising unnecessary primary capital, as it is cashflow positive and well funded. Accel and 01A participated alongside Salesforce Ventures and S32, following a prior tender and Series C valuation of $1.25 billion. Linear reports surpassing $100 million in annual recurring revenue, serving more than 40,000 paying companies and achieving 177% net revenue retention, with enterprise customers including Salesforce and Figma and broad adoption among AI companies such as OpenAI, Cursor, and Cognition. The company says its agent platform is installed in 95% of paid workspaces, with agents now creating half of work items, while agent-supported development activity has increased substantially in 2026. Linear plans to expand carefully across more than 30 open roles to support its broader product ambitions, while continuing employee equity refreshers and extended paid leave benefits.

Linear rebuilt its delta sync read path to help local-first clients efficiently catch up on changes after being offline, particularly in large workspaces that can generate nearly one million sync actions daily. Its previous Postgres-based approach became increasingly CPU-intensive and unpredictable because it had to scan large ID ranges while repeatedly filtering actions by user permissions and client subscriptions. The new architecture uses turbopuffer’s inverted indexes, which represent permissions and subscriptions as sorted posting lists of sync action IDs that can be unioned and intersected directly with requested ID ranges. Linear replicates sync-action metadata from Postgres to turbopuffer through a change-data-capture pipeline, then uses a two-stage process that filters lightweight metadata before retrieving full payloads from Postgres only for actions that will be sent to clients. To safely address replication lag, Postgres remains authoritative for a small recent portion of the log, overlapping with turbopuffer results that are deduplicated, while requests can fall back to Postgres when needed. Shadow testing verified that the new path returned the same actions as the existing system, and production results showed more stable p95 and p99 latency as workspaces, permission sets, and sync logs grew.

Linear describes building Linear Agent by allowing flexible, context-aware behavior within carefully designed boundaries rather than scripting fixed workflows. Its system prompt establishes communication standards, safety constraints, product-specific concepts, and default judgment rules, while its tools are designed to make appropriate actions intuitive and invalid ones difficult. Because Linear’s many specialized product operations require knowledge not inherent in general-purpose models, the agent uses “system skills” that package relevant instructions, metadata, and tools and are loaded selectively based on a task’s context. Linear intentionally avoids granting low-level SDK, CLI, or API access to reduce speculative or unsafe actions, accepting narrower capability in exchange for greater predictability. A custom underlying harness supports dynamic tool injection, context-sensitive approval for consequential actions, and asynchronous sub-agent work, giving the team control over efficiency, reliability, and emerging edge cases as agent capabilities evolve.