Agents
Agents
Agents is where you set up AI agents, put them in teams, run them on a schedule and see what they did and what it cost (module agents). An agent is a name, instructions and a reach: which modules' tools it may call. It always acts as a person, with that person's permissions and nothing more. The same page also reaches agent runtimes that live on your own machines.
In the sidebar under Agents → Agents. Talking to agents has its own page, Chat, under Home: see Chat. An agent needs a model to think with: add a connection under AI first.
What you see
The Agents page has these tabs:
- Chat: talk to an agent from here as well. The full chat page is Home → Chat.
- Playground: a code scratchpad. You write a JavaScript script that calls tools, run it as yourself without a model, and see what it returned, its console and its tool calls. Beside it, Reachable tools lists every tool you may call.
- Agents: the agents you own and the ones you may use. An agent another agent made sits beneath its maker. The filter offers All I can use, Mine, Shared with me and, for an administrator, Others' (as administrator). Switch the view to Agent teams to see your teams. Above the list you may see two cards:
- My assistant: your own private assistant, when your organisation has switched personal assistants on.
- Who looked at your agents: each time an administrator, or somebody allowed to read everyone's agent chats, read one of your agents' conversations or runs. It only shows when there is something to show.
- Runs: runs, newest first, with status, prompt, steps and start time. Click one for the full prompt, the output or error, the model cost, a Trace of the whole request and a Debug log of each model call.
- Calls: every tool call runs made, with its outcome (
allowed,deniedorfailed), the tool, the error, the argument names (never their values) and how long it took. - Triggers: standing instructions that run an agent on its own, with when, whether each is on, and how its last run went.
- Agent groups: named sets of your own agents that share model choices (see "Models for each job").
- Routing: which agent answers on which surface (Telegram, the web chat, the API), and a What would answer check.
- Runtimes: your machines and the agent runtimes on them.
- Budget: tokens used this UTC month against the monthly cap.
- Health: set-up problems the agent doctor finds, with a fix where one is offered, plus the size of the prompt and how many tools are offered.
An administrator sees everyone's runs and calls. Others see their own.
How agents work
Kinds of agent
- The general assistant. Chat with no agent picked. It has no instructions of its own and reaches everything you can.
- Agents you set up. A name, a description, instructions, a connection and model, a step limit and a reach. Each has one owner.
- Your own assistant. When an administrator switches on "every person gets their own assistant", each person gets one private assistant. You can rename it and set how often it speaks up on its own: Less (only when there is something to say), Normal (a short brief each morning) or More (alerts as they happen too).
- Agents made from templates. A template is a ready-made specialist, such as a bookkeeper or a researcher. An agent made from one gets the template's instructions and tool packs, never its maker's. When a template gets a newer version, you are offered the update once, and you can keep yours.
- Agents made by agents. An agent with May make agents switched on can make lasting agents of its own. Each holds no accounts, reaches no further than its maker, and waits for a person to approve it before it runs.
- Delegates and hand-offs. During a run an agent can ask another agent and relay its answer, hand the whole conversation to a specialist, hand it back, or pass it to its team's lead. A delegate never reaches further than the agent that asked.
- Runtime agents. Agents that live in a runtime on one of your machines (Hermes, OpenClaw, Claude Code or a custom one). Max sends them messages and work. They run there, with the reach that machine gave them.
Agents belong to one person
Agents are no longer shared. The Sharing button is now Give to: it hands an agent to somebody else, who owns it from then on. Where agents were shared before, an administrator sees Agents still shared with a preview of the move: each person who used a shared agent gets their own, keeping their chats and notes. Move now does it, and is safe to run again.
When somebody other than the owner changes your agent, you are told.
Settings on an agent
| Setting | What it does |
|---|---|
| Name, Description | How it is listed. The description is not sent to the model. |
| System prompt | Sent at the start of every run, before its documents and the conversation. |
| Connection, Model | Which model answers. Empty follows the model order. |
| Web access | Whether it may search the web. |
| Voice, Which voice | It answers out loud only when you sent a voice note, unless you asked otherwise. Needs a speaking connection under AI. |
| Step limit | The most tool-calling steps one run may take. |
| Tools shown per step | 8 to 128. Empty shows 24. The agent can search for the rest. |
| Recall memories | Each run starts with the person's most relevant memories and the agent's notebook. |
| Learns from its turns | After you correct it or ask it to remember something, it keeps a short lesson and tells you, with Undo. Skills it drafts wait for you. |
| May make agents | Lets it make lasting agents (see above). Only a person can switch this on. |
| Compact long runs, Compaction token threshold | Summarises older steps of a long run, keeping ids and amounts exact. |
| Conversation it reads | How much of a conversation it reads word for word, 1,000 to 60,000 tokens. Empty is 6,000. |
| Sandbox | Not built. Leave it on None. |
| Execution Profile | Standard (ReAct), Code-as-Action (writes one script that calls several tools), Minimal (Benchmark) or Plan (Planner) (plans only, changes nothing). |
Documents on an agent are extra texts, such as a voice.md, sent after its instructions in order. You can switch one off without deleting it.
Reach
What an agent may call is the smallest of these:
- what the person it acts for may do;
- the modules ticked under Modules (none ticked means everything that person can reach);
- its capability packs, named bundles of tools that only ever narrow;
- the outside MCP servers ticked under Outside tools.
Whether a call runs at once, asks a person first, or is refused is decided by the approval rules. What runs on its own on an agent's row lists each tool under Runs on its own, Refused and Asks first. Change a rule on Approvals.
Models for each job
Models on an agent sets a model for each job: text, writing, thinking, pictures in, documents, voice in, voice out and video, each with one backup. Use what it inherits takes the choice from its agent groups, then its team, then your default, then the workspace. You can set one job on many agents at once, with one Undo.
An agent group shares model choices between your agents. When an agent is in two groups, the lower priority number wins. A group does not make agents work together: that is a team.
Teams
An agent team has a lead and members. You talk to the lead; it hands parts of the work to members, and each hand-off is a task in Tasks. Each member reports to the lead or to another member.
- New agent team makes one by hand. New agent team from template makes one from a use case (shop, solo, agency, trade, personal, content, infra or money). Members you lack access for show Needs access: press Ask admin, and later Add the missing members.
- A team's Work panel shows each member's role card (its job in one line and the tools it may use when handed work), its open tasks and what it cost this month. Limits set what one request to the lead may cost in AI (default $0.50) and how long the lead waits on a member (default 30 minutes).
- Make it the front door makes an agent the one that first reads requests and routes them to one agent or a team. An administrator can Undo front door, which puts it back exactly as it was.
Runs
A run is one request to an agent and everything it did. Its status is queued, running, awaiting_approval, done, failed, stopped or merged (folded into an older queued message). The step limit caps tool calls. The same failing call repeated several times in a row stops the run.
Stop ends a run after the step it is on. Pause is the stop switch for a whole agent: every run going, waiting or queued stops and its cards are withdrawn. Only a person can start it again.
Conversations
A chat is kept, so a reload, a second tab or Telegram all show the same thread. An agent's and a tool's turns are removed 30 days after the conversation's last activity; your own words are kept for 12 months, so agents can recall your past chats.
Triggers
A trigger runs an agent's prompt on its own: Every so many minutes, Once a day, at a time, or A number of times a day, at random times between hours you set. It runs as whoever created it. What an unattended run writes to a project's memory is proposed for a person to accept. Only the creator or an administrator can delete it or change its prompt. An agent can change its own trigger when you ask it to in chat.
Budget
A monthly token cap for the whole workspace, per UTC month. Empty is unlimited and 0 stops every run. Warn at (%) sets when Health starts warning. When AI spend is installed, the cap is kept there.
Runtimes and hosts
A machine (host) is somewhere agent runtimes live. Max reaches it over your private mesh. A host nobody has been granted to is open to everyone here; one with members is seen only by them and administrators. A runtime is an agent process on a machine. Its Kind decides what it can do: chat, long runs, its own scheduled jobs. Its key is encrypted and never shown again. May act on the host itself lets it work outside its own workspace; leave it off unless you would trust it with a shell.
What you can do
Create an agent
- Open the Agents tab and press New agent, or New from template to start from a template.
- Fill in Name, Description and System prompt, and pick a Connection or leave it empty.
- Press Create. If the template suggests approval rules, tick the ones you want and press Make the rules I ticked.
- Press Run on its row to try it.
Limit what an agent may reach
- Press Modules on its row and tick the modules it may use. Save.
- Press Outside tools to tick outside MCP servers.
- Press What runs on its own to see the result for each tool.
Give an agent to somebody else
Press Give to on its row and pick the person. They own it from then on.
Make a team
- On Agents, switch the view to Agent teams and press New agent team (or New agent team from template).
- Name it, pick the Lead, tick each member and choose whom it reports to. Press Create.
Put an agent on a schedule
- Open Triggers and press New trigger.
- Pick the Agent, give it a Name and a Prompt, and choose How often.
- Press Create. Change time changes the schedule later and keeps everything else.
Choose which agent answers where
- Open Routing and press Bind an agent.
- Pick the Surface and the Agent. Leave Target blank for the whole surface, or name one chat. A chat binding wins.
- Use What would answer to check.
Set a monthly token cap
Open Budget, press Set cap, enter Monthly tokens and Warn at (%), and save.
Add a machine and a runtime
- Open Runtimes and add a machine: Name, Mesh address, Operating system and Location.
- Press Add runtime: pick the Host, name it, choose its Kind, and enter its Address and API key.
- Press Test, Sync agents and Ask.
Read what an agent learned
Press Learned or Notebook on its row. Correct or remove entries under Knowledge → Agents.
Who can do what
| Permission | Label | Lets you |
|---|---|---|
agents:read | View agents and their runs | Open the page and see agents, your runs and calls, triggers, routing, runtimes, the budget and Health. |
agents:write | Create and run agents | Create, change, give away and delete your agents; chat and run; make triggers, teams, groups and bindings; ask runtimes and start their work. |
agents:manage | Register agent runtimes and change what they may reach | Add and remove machines and runtimes, change their settings and jobs, run their tools directly, and switch personal assistants on or off. |
agents:assistant | Use your own assistant | See and rename your own assistant, set how often it speaks up, and let support read its chats for a while. Everyone holds this. |
agents:audit | Read everybody's agent conversations and runs | Read only. Each read is kept in the audit trail and shows to the agent's owner under Who looked at your agents. |
agents:manage-others | Make and change other people's agents and teams | Lent to an administrator's own API key for a while. Each use is approved by a person. |
Through an API key, some changes that would widen what an agent reaches are refused, and the tool says to sign in instead. See People, roles and permissions.
What happens on its own
| Job | When | What it does |
|---|---|---|
triggers | Every minute | Runs every switched-on trigger that is due, as the person who created it. |
drain-queues | Every minute | Picks up conversations with messages waiting and nobody working on them, for example after a restart. |
past-chats-embed | Every 5 minutes | Indexes your own chat messages so agents can recall them by meaning, when an embedding model is connected. |
usage-legacy-copy | Every 10 minutes | Copies the older usage log into the record AI spend reads. |
personal-assistants | Every 10 minutes | Where personal assistants are switched on, makes one for each person who has none, and stops or restarts them as needed. |
style-review | Every hour, at 25 past | For each person whose own clock is between 03:00 and 03:59, reviews their style card once for the day just gone. No model call for a day without chats. |
consolidate-conversations | Every day at 03:15 UTC | Summarises conversations idle for a day and proposes what is worth keeping to their people and projects. |
routing-retention | Every day at 03:20 UTC | Deletes front-door routing records older than 90 days. |
prune-sessions | Every day at 04:00 UTC | Removes agent and tool turns after 30 days and your own words after 12 months. |
run-event-tenancy | Every day at 04:30 UTC | Checks that every recorded run event belongs to its workspace and logs a warning if not. Changes nothing. |
prune-debug | Every day at 04:45 UTC | Deletes run debug entries past the workspace's retention, and timing rows past at least 7 days. |
Needs
Nothing else switched on. A model connection under AI to think with; Approvals to ask before acting; Knowledge for memories; Skills for skills.
What agents can do
Each tool runs under the caller's own permissions. Read-only tools run at once. Every other tool asks a person first under the default rules, unless a rule on Approvals says otherwise. Tools marked deletes remove something; outside acts on another machine. Over MCP the tools are named agents__list and so on.
Inside a run
agents.delegate: ask another agent and get its answer back.agents.handTo,agents.handBack: hand the conversation to a specialist, or give it back.agents.escalate: pass the conversation to the team's lead.agents.findTools(read-only): search every tool the agent may use.agents.code_action: run a script that calls several tools in one step.agents.recall(read-only): the person's past chats on every surface, last 12 months.agents.ask: ask one of your agents something and get its answer.agents.research: deep research in the background, then an answer.
Agents, documents and packs
- Read-only:
agents.list,agents.get,agents.docs,agents.packs,agents.children,agents.shares - Change:
agents.create,agents.update,agents.remove(deletes),agents.addDoc,agents.updateDoc,agents.removeDoc(deletes),agents.createPack,agents.updatePack,agents.clearUnknownPacks,agents.removePack(deletes),agents.review(let an agent another agent made run, or switch it off),agents.setOwner(give it to somebody; administrators only) - Sharing:
agents.share,agents.replaceRoleShare,agents.unshare. Sharing is switched off now:agents.shareis refused.
Models and groups
- Read-only:
agents.models,agents.groups - Change:
agents.setModels,agents.undoModels,agents.createGroup,agents.updateGroup,agents.removeGroup(deletes),agents.setGroupModels
Runs, cost and health
- Read-only:
agents.runs,agents.runDetail,agents.runTrace,agents.runDebug,agents.timings,agents.usage,agents.activity,agents.budget,agents.doctor,agents.toolDecisions,agents.promptSize,agents.routing - Change:
agents.stop,agents.pause,agents.unpause(a person decides)
Routing and the front door
- Read-only:
agents.bindings,agents.bindingTargets - Change:
agents.bind,agents.unbind(deletes),agents.becomeFrontDoor,agents.undoFrontDoor
Schedules
- Read-only:
agents.schedules,agents.myTriggers - Change:
agents.schedule,agents.reschedule,agents.unschedule(deletes),agents.updateMyTrigger
Conversations and learning
- Read-only:
agents.conversations,agents.conversation,agents.learnings - Change:
agents.renameConversation,agents.removeConversation(deletes),agents.undoLearning
Personal assistants
- Read-only:
agents.myAssistant - Change:
agents.setMyAssistant,agents.supportWindow(let support read your assistant's chats for a while),agents.personalAssistants(switch on or off for the organisation),agents.sweepAssistants,agents.adoptPersonal
Teams and templates
- Read-only:
agents.teams,agents.teamTemplates,agents.teamWork,agents.specialists - Change:
agents.createTeam,agents.renameTeam,agents.removeTeam(agents are kept),agents.moveToTeam,agents.setTeamLead,agents.setReportsTo,agents.setRoleCard,agents.setTeamLimits,agents.teamFromTemplate,agents.completeTeam,agents.fromTemplate,agents.requestTemplate,agents.applySet,agents.linkTemplate(administrators only),agents.undo(remove an agent made from a template)
Hosts and runtimes
- Read-only:
agents.hosts,agents.hostMembers,agents.hostJob,agents.runtimes,agents.testRuntime,agents.runtimeAgents,agents.runtimeRuns,agents.refreshRuntimeRun,agents.runtimeJobs,agents.runtimeLinks - Change:
agents.updateHost,agents.removeHost(deletes),agents.removeHostMember(deletes),agents.updateRuntime,agents.removeRuntime(deletes),agents.updateRuntimeSettings,agents.restartRuntime(outside),agents.deprovisionRuntime(deletes; uninstalls a Hermes runtime Max installed, cannot be undone),agents.syncRuntimeAgents,agents.addRuntimeAgent,agents.updateRuntimeAgent,agents.removeRuntimeAgent(deletes) - Work on runtimes (outside):
agents.askRuntime,agents.broadcastRuntimes,agents.startRuntimeRun,agents.stopRuntimeRun,agents.invokeRuntimeTool - Runtime jobs and links:
agents.addRuntimeJob,agents.updateRuntimeJob,agents.pauseRuntimeJob,agents.resumeRuntimeJob,agents.runRuntimeJobNow(all outside),agents.removeRuntimeJob(deletes),agents.addRuntimeLink,agents.removeRuntimeLink(deletes)
Granting a host to someone, and its mesh address, stay with a person.
Related
- Chat
- AI: the model connections agents think with.
- Skills, Knowledge and Connectors: what agents load, remember and act as.
- Approvals: where waiting calls are decided.
- Agents and approvals
- API keys and MCP access