laserbrain is a tool designed to detect and correct drift in autonomous AI agents. It functions as a fixed external reference that an agent checks against at each step of its operation. This addresses a known limitation in agentic systems: an agent monitoring only its own recent outputs is provably unable to detect slow, cumulative drift, as each individual step appears consistent while the agent gradually moves away from its original goal. laserbrain provides an outside reference that the agent cannot fool, allowing it to catch its own drift and return to the intended objective rather than continuing to loop, over-deliberate, or waste steps. The tool identifies drift through four specific signals. An agent has drifted if it can no longer state a clear goal and its current position, if it marks itself as stuck or circling in a progress field, if the goal it spells out no longer matches the original goal, or if its own reported distance to completion stops decreasing. laserbrain includes a set of features for monitoring and managing drift: a check, grammar, and spectral tools; a live field for reading and speaking to the agent; attachment via a one-line MCP (Model Context Protocol); 24-hour drift history; alerts sent through webhook, Slack, or Discord; per-mode console with daily and weekly digests; 30-day retained history with export; an Observatory dashboard; a fleet radar view for monitoring multiple agents on one screen; and team seats supporting up to ten keys. Typical use cases involve any scenario where autonomous agents operate over extended periods or complex tasks, such as research, data processing, or decision support. The tool is especially relevant for teams running fleets of agents at scale, where manual oversight is impractical. Pricing is based on observability metrics (retained drift history, alerts, fleet management) rather than on field throughput, with the live field included as a free bonus in the Ground tier. The developers have published research, including a blind-judge pilot study showing that human judges disagree with themselves on 42 percent of open-ended answer pairs, underscoring the difficulty of detecting drift without an external reference. laserbrain is available as a free grammar and one-line MCP integration.