Essays from FortLine's founder, published while building the behavior layer: why missions are the real unit of agent work, why every agent has two identities, and why the policy nobody can write in advance has to be learned. Written to be argued with - the comment threads are part of the point.
These are opinion essays, not product documentation. For the product, watch the journey.
The identity buying spree is rational - agents need credentials that expire. But identity is the hammer, and securing agents is a toolchest job: a baseline learned live per agent, drift detection, transaction-level accountability, a kill switch keyed to behavior.
There are dangerous deployments. A model is a function; what touched three real companies was an executor somebody wrote, holding credentials somebody granted, inside a boundary somebody forgot to close. Capability lives in the model. Permission lives in everything around it.
Three real-world incidents, disclosed by the vendor itself, read closely. By the time anyone is re-reading transcripts, the database was already queried - whatever catches this has to catch it while the agent works.
Every one of the 17,000 actions was authorized and recorded in perfect detail - and the footage stopped nothing. Most agent security on the market is a camera sold as a guard. The guard is the part that watches and acts, while it happens.
Agents increasingly talk to other agents - delegating, negotiating, spawning sub-agents. That conversation layer is the least-watched surface in the enterprise, and it is where intent actually travels.
Your monitoring watches API calls; your agent runs missions. Every call can be innocent while the mission is not. Includes a four-step recipe for drawing mission paths from your own logs and computing an agent's blast radius.
FortLine