At a technical level, an AI agent is a large language model that repeatedly calls itself. It uses provided tools to gather information and build its own context until it has enough data to achieve a predefined goal, making complex tasks autonomous.
Directly giving an agent a full API schema causes hallucinations. Linear solved this by developing a "skills" architecture where the agent dynamically loads specific, pre-defined toolsets and instructions (e.g., "create an issue") based on the user's task.
Counterintuitively, giving an AI agent too much upfront context or overly specific instructions degrades its performance. The best practice is to provide minimal guidance and instead equip the agent with tools to dynamically fetch the necessary information itself.
A chatbot is a necessary interface for multi-turn interactions but shouldn't be the primary entry point. The most effective domain-specific agents are accessible from natural "on-ramps" within a user's existing workflow, such as a Slack conversation or a meeting summary.
Instead of using one large model for all tasks, Linear employs a small router model. For high-frequency use cases like creating issues, it routes the request to a smaller, highly-optimized model and prompt, saving costs while improving performance and reliability.
When evaluating AI agents, it's crucial to distinguish between areas that require strict consistency (e.g., setting a ticket status correctly) and those that benefit from natural variance. Over-evaluating for consistency in creative responses can generate false negatives and limit the agent's utility.
To understand how users would naturally interact with their agent, Linear quietly replaced its procedural Slack bot with the new AI. This revealed unexpected behaviors, like users simply typing "@linear do the right thing," providing invaluable real-world usage data.
The primary bottleneck in applied AI is not model capability but human integration. There is a massive "capacity overhang" where models are far more powerful than how we currently apply them in daily workflows. The focus should be on better application, not just better models.
When building a new agent feature, the best practice is to use the most powerful model available to explore the full capability space. Only after establishing clear success criteria and evaluation frameworks should you optimize by using smaller, more cost-effective models.
A generic API lets users build their own processes, which can be inconsistent. By building a native agent, a company like Linear can embed its specific philosophy and best practices (e.g., "the Linear method") into the agent's core behavior, ensuring quality.