Govern Long-Running Agents
Long-running agents need owners, limits and a way to intervene. Lightspeed gives teams practical controls for work that lasts beyond a single conversation.
A long-running agent needs the same things as any production system: an owner, limits, a way to intervene, and a record of its work. An agent investigating incidents for months needs more supervision than a conversation someone can simply close.
Lightspeed makes that supervision part of the way agents run. Sessions are durable workflows, so their progress and controls remain available across worker restarts.
Someone owns the work
Company sign-in and roles determine who can use an agent, configure its tools, or administer the team. Profiles give operators a reusable setup for each job.
An existing business workflow can take responsibility for the agent’s work too. It starts a managed session, supplies the operations the agent may request, and decides what to do with the result. Your incident or onboarding process stays in charge as the agent handles one step within it.
Delegation has limits
A planner can start builders, testers and reviewers, each using an approved profile. Limits bound how many agents it can create, how many can work at once, and how long delegated work may run. Recurring bots also have a daily run budget.
This makes delegation a capacity decision operators can understand before the work starts.
Operators can intervene
Operators can steer active work with a new instruction, queue another task, or cancel a run. These controls remain available while the agent is waiting on a model or tool.
For consequential actions, a business workflow can hold the decision at an approval step. Stopping an investigation and authorising a production change are then separate, deliberate operations.
Waiting has a deadline
A task that depends on a reply needs a time limit. An incident workflow can give an investigation fifteen minutes, then escalate if it has no useful answer. Workflow-backed tools support durable waits and deadlines, so these policies can live in the process rather than in a reminder to the model.
The work stays visible
Sessions link to their sub-agents, and bot activity connects incoming events to the conversations that handled them. A reviewer can follow who requested the work, what the agent called, and which results it received. The audit article explains that record in more detail.
Prompts, tools and models change over the life of an agent. Lightspeed’s evaluation harness lets teams exercise representative tasks and check the results before rolling out a change.
The session’s progress and control decisions are recorded in durable state. Another worker can recover that state and continue the work. The team can therefore supervise a process lasting weeks without depending on one machine staying alive for the whole task.
Read next: Secure Agent Fleets. For setup, see agent and tool access and workflow integration.