GTM Agents · Capability
Human approval for GTM agent actions
Human approval means a GTM agent pauses for explicit sign-off before a sensitive action — a CRM write, an outbound touch — and resumes only when a person responds. Lumitra combines approval gates with tool-level allowlists and execution budgets, so agents get autonomy for the work and humans keep the veto.
An agent acting on revenue systems should ask first
The value of a GTM agent is that it acts — researches an account, updates the CRM, routes a lead, drafts the follow-up. The risk of a GTM agent is exactly the same thing. An autonomous system that can write to your CRM and touch your prospects without a human in the loop is not an execution advantage; it’s an incident waiting for a timestamp. The category-defining question is not “can it act?” but “who decides when it may?”
Approval gates
Sensitive actions pause for explicit approval before the write continues — the agent prepares everything, a person releases it.
Tool allowlists
Each agent gets only the tools and MCP actions its workflow needs, at server and tool level — an agent that can’t reach a system can’t misuse it.
Bounded execution
Work-time and tool-step budgets, no-op and repeat-loop detection, and bounded delegation keep multi-step runs inside limits you set.
How an approval-gated run works
- The agent works autonomously — planning, researching, and using its allowed tools through multi-step continuation, with durable memory and persisted results carrying context.
- It reaches a sensitive action — and pauses with an explicit approval request describing what it wants to do.
- A human responds — approve and the write continues; decline and the agent moves on without it. Either way, the run stays inside its budgets.
- Everything is recorded — the request, the decision, the tool calls around it, and the structured completion report land in the audit trail.
Frequently asked questions
Which actions require approval?
Sensitive ones — actions that write to systems of record or reach outside the workspace. An agent can research, read, and prepare autonomously, then pause for explicit approval before the sensitive write continues. The run resumes only after a human responds, and the request, response, and outcome stay in the session record.
Is approval the only control?
No — it's the last of several layers. Each agent carries only the built-in tools and MCP actions its workflow needs, through server- and tool-level allowlists. Job-mode runs enforce work-time and tool-step budgets and detect no-op or repeat loops. Delegated child sessions are bounded and report completion, failure, or timeout to the parent.
Doesn't approval defeat the point of autonomy?
Autonomy and control are not opposites — the expensive part of GTM work is the research, context assembly, and preparation, which agents do without interruption. Approval applies at the moments where a mistake is costly: the CRM write, the outbound message, the record change. You keep the leverage and the veto.
Can approval requirements differ per agent?
Yes. Agents are configured individually — interactive or autonomous job mode, tools, variables, model settings, and execution limits — and configuration is versioned, so changing an agent's controls doesn't silently rewrite the setup that produced older runs.
How do I see what was approved after the fact?
Every approval is part of the session's audit trail, alongside the tool calls and outputs around it. You can inspect what the agent asked, who approved it, and what happened next — see audit trails for the full picture.
Watch an agent ask permission
Book a demo and see a live agent pause on a CRM write, wait for approval, and log the whole exchange.