Agent Lab
Meet your workspace's agents, read what each can do, talk to them, create a new colleague, and teach agents with feedback.
Agent Lab is where you meet, create and configure your workspace's agents: a governed company, not a catalog. Each agent has a persona, a model, a set of capabilities and a permission role. This page shows you how to read the roster, talk to an agent, create a new colleague and teach agents with feedback.
Before you start
- Open Build › Agent Lab.
- Everyone can read every agent, talk to them and give them work. Creating an agent needs the right to create agents, or the admin role.
- Axiom is the Lab's resident. Click Ask Axiom about the roster or how to shape a new colleague.
Meet the agents
Open Build › Agent Lab.
The roster is grouped: Buddies, Internal agents by department, Integrations for external agents wired in, Other, and Test & retired.
Search with Search by name, specialty or description….
Click an agent to open its page.
Each card shows the agent, its specialty and whether it is online. An external assistant is marked external connector.
For owners and admins, a switch at the top chooses the user view or the admin view. Your choice follows you onto an agent's page.
Read what an agent can do
In the user view, an agent's page has two parts.
Read What it can do: every permission the agent holds, in plain words, and how many are available now.
Pick a line to see what it means for that agent, whether it leaves a record, and Tools this permission brings.
Read When it steps in: its specialty, the work that is the agent's, and the work that is not.
The Know your agent card counts how many capabilities you have read. Start again clears only your own notes.
In the admin view, the page also has Character, with persona, model, behavior, limits per turn and memory, and the Orchestration pipeline.
Talk to an agent
On the agent's page, click Chat.
The agent answers as itself, within its own role. The header shows its permission scope.
Type your message.
To turn the conversation into a task, click Create task from conversation.
History lists your past chats; New chat starts a fresh one.
An external connector cannot be chatted with here: it acts through its own app. Manage it in Settings › Your external agent.
Create a new agent
Click Create new agent.
The page A new colleague asks you to describe the colleague you need.
Write it in I need someone who…, the way you would tell a person. A sentence is enough; a paragraph is plenty. Or pick one of the examples.
Click Read it.
The page shows Nexa is reading. with each stage. You can stay or leave: the proposal will be waiting in Agent Lab.
Read the proposal:
Review What … will be able to do, grouped as Available, Needs your choice and Not available.
For each line under Needs your choice, name what it may use, for example which websites it may reach. Nexa never picks these on its own.
Read In the description, not given: what you asked for that the colleague will not have, each with the reason. Nothing is dropped in silence.
Optionally, under Check the short text, add a line the colleague reads in every conversation, for example "always ask before sharing a price". Click Add this line.
Try it before you bring it in: under Give the candidate a task before you decide, type a task and click Ask the candidate.
The candidate answers within about 45 seconds and shows WHAT THE CANDIDATE REACHED FOR. Nothing is saved.
If it is not right, click Rewrite the description, or Throw away.
Click Bring … in.
The page says the colleague is on the team, with Say hello to ….
When a permission needs your admin
If a capability needs an admin's word, accepting sends a request. The page says the colleague is waiting for your admin, and nothing is created yet. When the admin grants it, the draft reads YOUR ADMIN ANSWERED: click Accept again. The proposal is kept, and nothing is read anew.
Revise an agent
This is for agents your workspace created.
On the agent's page, click Revise.
Describe the agent as it should be from now on, starting from today's introduction. Click Read it.
The agent stays as it is until you accept.
Review the proposal and click Revise ….
The switches that decide what kind of agent it is, for example whether it runs projects, are fixed when it is created.
Check that an agent holds what its persona asks for
In the admin view:
Open Capabilities, then Persona ↔ permission consistency.
Click Compile persona. Nothing is saved; it can take about 30 seconds.
The result says the agent holds everything its persona asks for, or how many things are missing.
The same tab shows Operating frames, What it can use, What wakes this agent, with Revoke, and Fine-grained access, including the sites Web search may reach. Empty lists mean every site can be found and read.
Pause an agent's autonomy
On the agent's page, use the Autonomy on control.
Read the dialog: the agent stops what it starts on its own, such as scheduled runs, event triggers and automatic wakes. Project work, delegations and chat continue.
Click Pause …, or Keep running.
Delete an agent
On the agent's page, click Delete agent.
Click Confirm delete.
Teach an agent with feedback
Under an agent's answer, click the small face icon, Tell … what went wrong.
Write in What should have been different? Be specific: what missed the mark, and what would have been right.
Choose who sees it:
Click Send to ….
The page confirms that the agent will remember this.
The agent's page shows Learning from experience with how many lessons it has. The Experience tab lists the most effective ones.
What you will see
Good to know
- The read has no time limit, and leaving the page does not stop it.
- The trial task runs on the real candidate and saves nothing.
- Revise and Delete appear only to those who may edit or delete that agent, and never on Nexa's own agents.
- When an agent proposes creating or reconfiguring an agent, the proposal goes to Pending Actions for a person to approve.
- Pausing autonomy stops only what the agent starts on its own.