Lifecycle

How do I manage the lifecycle of agents?

An HR lead and an IT security colleague review a checklist as a small AI agent walks with them from a test workshop into a busy office

Anyone should be able to try an idea with an agent. Not every experiment should touch real customers, real systems or the budget your live agents depend on.

The challenge

When experiments and real work live side by side, things go wrong quietly. A test agent gets access to the live CRM. A working agent is changed on a Friday afternoon without anyone noticing. Experiments use up the credits that business-critical agents need. Banning experiments is not the answer. Separating them is.

A short story

Emma in HR builds an onboarding agent. Like every new agent, it starts in the sandbox, connected to a test version of the HR system.

She prepares the real HR connection in advance. It just waits, marked "Activates in production".

When the agent is ready, Emma clicks Request promotion. IT security and her team lead review it and approve. The agent goes live with the real connection, without being rebuilt.

Two months later Emma changes one of its protected documents. An automatic check flags the change and a quick re-review starts, while the agent keeps working.

Emma could move fast. IT could sleep at night.

What does the journey look like?

Two phases, one clear gate between them.

  1. Step 1SandboxBuild and test with test data and test keys.
  2. Step 2ReviewChosen reviewers approve or reject.
  3. Step 3ProductionReal work, real connections.
  4. Step 4Re-reviewAfter key changes, or on a schedule.

How does Abundly manage the agent lifecycle?

  • 1

    Keep experiments safe

    Free to try, with no access to what matters.

    • Sandbox and production phases: every new agent starts in the sandbox.
    • Mark sensitive capabilities, keys and connections as production only. Sandbox agents can use test versions instead.
    • Sandbox agents only talk to other sandbox agents.
  • 2

    Go live with a review

    A clear, recorded decision before real work starts.

    • Promotion reviews: owners request promotion, and reviewers you choose approve or reject.
    • Prepare production access in the sandbox, so it switches on the moment the agent is approved.
  • 3

    Keep live agents in check

    Changes to working agents do not slip by.

    • For each kind of change, choose to allow it, check it or block it in production.
    • Only documents you have protected are watched, so protect the ones that matter.
    • Flagged changes lead to a re-review while the agent keeps running, or move it back to the sandbox.
    • Optional regular re-reviews, even when nothing has changed.
  • 4

    Protect the budget for live work

    Experiments never starve real work.

    • Reserve credits for production agents. If the balance drops to that level, sandbox agents pause first.
    • Filter Usage to production agents only, to see how many credits your live work uses.

How do I manage many agents at once?

  • 5

    One overview, bulk actions

    For admins with dozens or hundreds of agents.

    • Agent management shows each agent’s phase and status, such as "Promotion pending" or "Re-review pending".
    • Move several agents between phases in one go, or let a Workspace Manager agent do it for you.
  • 6

    People decide, not agents

    The gate stays in human hands.

    • An agent can never change its own phase, and approvals are always made by people.
    • Every phase change is recorded in the audit log.

Where do I start?

  1. 1

    Turn it on

    Agent lifecycle management is part of the Enterprise plan. Abundly enables it for your workspace, then your workspace admin switches it on in workspace settings.

  2. 2

    Mark what is sensitive

    Make your real systems and keys production only, and reserve credits for live agents.

  3. 3

    Choose your gate

    Decide who reviews promotions and which changes to live agents need a second look.

Good to know

Lifecycle phases are part of the Enterprise plan. Nothing changes for running agents until you mark resources as production only.

See it in your own workspace

We will walk you through the platform with your own agents and rules in mind.

Request a demoMore questions answered