Managing Your Agents

An agent isn’t a one-and-done tool. It builds up memory over time, runs more and more on its own, and quietly racks up usage. This page covers what you need to know to keep an agent healthy long-term — pausing, cleaning up memory, checking costs, tidying automations, managing your skill library, adjusting its computer specs, transferring ownership, and deleting.

Fast hire — create a new employee instantly, no conversation needed

Press “Hire” and a new hire shows up right away — there’s no chatting with a manager first to decide what to assign. Starting an employee from scratch goes through this button and nothing else (installing a preset from the Market, which brings over an employee someone already built, is a separate path). A confirmation dialog appears; click “Confirm” once, and a hire with no name or role yet appears immediately, asking on its own: “Tell me what you’d like me to take care of.”

Through the conversation that follows (the onboarding interview), the new hire fills in its own name, role, and personality based on how you answer.

This button is available on web, iOS, and Android, though it sits in a different place and reads differently — on web it’s + Hire at the top of the Office screen and the confirm button says Confirm; in the mobile apps it’s Hire under the + menu at the top right (or the empty-state button when you have no employees yet) and the confirm button says Hire.

Hiring straight from a team’s screen is web only.

A hiring suggestion card may show up too

The hire confirmation dialog sometimes carries a hiring suggestion card. When one of your current employees is clearly overloaded — a backlog of pending work, a recent run of requests concentrated on that one employee, or a heavy load of recurring schedules — UpServe surfaces that evidence on a card and asks “How about hiring someone to share this employee’s load?”

Accept it and the card’s “Hire someone to help” button hires right away, carrying “why this hire happened” into the new hire’s onboarding interview.
Not interested? “Not now” dismisses it, and that same evidence won’t be suggested again.
No card showing is completely fine. It’s a side note, not a requirement — with no card, the plain “Confirm” button hires just the same. (When a suggestion is shown, that button is relabeled “Hire directly” — meaning you’re bringing someone new on regardless of the suggestion.)

This card works the same way on web, iOS, and Android.

What’s different during onboarding

While the new hire is editing its own profile (name, role description, personality) based on your answers, you won’t see the usual approval prompt that shows up before risky actions. Instead, a change card appears in the chat. It shows the before/after content side by side, and if you don’t like the change, the card’s “Revert” button restores the previous content right away.

Once the new hire summarizes its proposed name, mission, and description and asks you to confirm, and you agree, onboarding ends, and you can use the rest of this page’s management features (active/inactive, memory cleanup, logs, and so on) exactly like any other agent.

Onboarding doesn’t end on its own just because time passes — leave the conversation mid-way and the new hire just waits. If it sits untouched for a long while (72+ hours), the progress card at the top of the chat switches to “Stalled.” Either way, just talk to the new hire to pick it back up, or use the agent page → Settings subpage → “Resume Onboarding” button.

At a glance

What you want to do	Where to do it
Hire a new employee instantly, no conversation	Web: Office screen → ”+ Hire” · Mobile: ”+” at top right → “Hire”
Hide an agent from the list and block all runs	Agent page → `Settings` subpage → “Agent Status” toggle → INACTIVE
Pause automatic runs temporarily	Agent page → `Schedules` subpage → toggle each schedule OFF
Stop the agent from waking up on its own	Same place, toggle the `Heartbeat` schedule OFF
Undo something the agent learned wrong	Ask the agent in chat: “Forget that note about…” / refresh the Brain panel
Wipe all memory and conversations	Agent page → `Settings` subpage → “Initialize Agent”
See how many SU you used this month	Agent page → `Logs` subpage
Rate whether an automatic run’s result was useful	Right below that result message — 👍/👎
Clean up unused or outdated skills	Skills (`/skills`) page → “My Library” tab → cleanup banner/modal
Change the specs of the computer it runs on	Agent page → `Settings` subpage → “Assigned PC Specs”
Transfer a personal agent to an organization	Office board → drag the agent card onto an org section (or an org team)
Move an org-owned agent back to your personal ownership	Agent list ”⋯” menu → “Transfer ownership (to member)” (org OWNER only)
Remove an agent you no longer need	Agent page → `Settings` subpage → “Delete”

1. Active / Inactive — lock an agent without deleting it

If you want to put an agent on hold without removing it, flip the Agent Status toggle to INACTIVE.

On the agent page → Settings subpage, find the “Agent Status” row and switch it OFF (INACTIVE). Here’s what happens:

The agent is shown as inactive in your list (dimmed or marked separately from active agents).
All automatic runs — schedules, Heartbeat, webhooks — are blocked.
However, you can still chat with it directly. (If it was auto-paused for inactivity, sending a message reactivates it automatically.)

Switch it back ON (ACTIVE) at any time to restore everything immediately. Settings, memory, and skills are fully preserved.

This toggle is not shown for system agents.

Away from UpServe for a while? Your agents pause themselves

Even without touching anything, if you haven’t used UpServe at all for about a week, your agents are automatically switched to INACTIVE. You’ll get a heads-up notification one day before that happens, and another notification when the switch actually occurs.

The key thing to know: “using UpServe” counts across any of your agents. You don’t have to open a specific agent — as long as you chat with any agent or open the app, all of your agents stay active. This means a background autonomous agent quietly doing its job doesn’t get shut off just because you rarely open its own page. It only pauses if you’ve stepped away from the service entirely.

Waking it back up is simple. An agent auto-paused this way returns to ACTIVE the moment you send it a message — no toggle to flip.

This automatic switch only applies when you’ve been away from the service entirely. If you turned the toggle off yourself, or the agent was paused due to low SU, sending a message won’t turn it back on — fix the underlying cause first, then turn it on manually.

2. Pausing — turn off autonomous runs for a while

UpServe agents can wake up without you in three ways:

Schedules — fire at a set time or interval (e.g., every morning at 9)
Heartbeat (autonomous activity) — wakes the agent on its own, self-paced, during defined active hours
Webhooks — external services ping the agent

If you want an agent to sit still for a few days, switch these triggers off. Each is a single toggle, and turning it back on restores the same setup.

Turn off schedules

Open the agent page and switch to the Schedules subpage (/agents/<agent>/schedules).
Flip the toggle on the right side of each schedule card to OFF.
When you’re ready, toggle it back ON and it resumes from the next scheduled time.

You don’t have to delete schedules to take a break — toggling is the easiest way.

Turn off Heartbeat (autonomous activity)

Heartbeat is the feature that wakes the agent inside defined active hours so it can ask “anything to do right now?”. The gap between check-ins isn’t fixed — it adapts to activity, stretching when things are quiet and tightening when the agent is busy. There’s one way to disable it:

Quick OFF: on the Schedules subpage, toggle the Heartbeat card OFF. (The editor always keeps at least one active-hours band, so emptying all bands to disable Heartbeat isn’t supported.)

While Heartbeat is off, the agent only runs when you message it directly or when a schedule or webhook fires.

Turn off webhooks

On the Webhooks subpage, switch the webhook OFF and external pings won’t wake the agent.

If you just want to stop the current burst, you can also chat “stop for today.” But that only applies to the run you’re in — the next scheduled time will still wake it. Use the toggles above for a real pause.

3. Cleaning up memory — undo what the agent learned wrong

Every agent keeps a personal notebook (Brain) of what it has learned: your preferences, frequently-used data, lessons from past tasks. Sometimes a note ends up wrong or outdated.

See what’s in the Brain panel

In the agent chat view, click the brain icon in the right-side toolbar to open the Brain panel. You’ll see:

SOUL.md / AGENT.md / MEMORY.md — the agent’s identity, operational guidelines, and memory index
User notes (user) — your preferences and context
Feedback (feedback) — lessons captured automatically after tasks
Project (project) — info on ongoing projects
Reference (reference) — frequently-consulted source material
Other (other) — memory files that don’t fit a specific type

Click any file to read or edit it. If the panel looks stale, hit the refresh button in the top-right of the panel.

Fixing wrong notes

The easiest way is to just tell the agent in chat. You don’t have to edit files by hand.

Situation	What to say
Drop a specific note	”Forget the note about me liking spicy food.”
A fact changed	”My office address changed — please update memory.”
Refresh how the agent does something	”We don’t use that report format anymore. Clean up the related feedback notes.”
Start completely fresh	See “Full reset” below.

Full reset — wipe conversations and memory together

On the agent page → Settings subpage → the danger zone at the bottom, click “Initialize Agent”. The following are deleted:

All conversation history
The entire Brain (notes and memory)
Canvas documents and in-progress tasks
Ad-hoc schedules (Heartbeat and Sensor Watch schedules are kept)
Attachments and screenshots
The agent’s sandbox (its code execution environment) — a fresh one is created next time

What’s kept: the agent itself (name, tools, skills, config), the Heartbeat schedule, and any Sensor Watch schedules. Sensors are treated as infrastructure rather than conversation state, so resetting the agent doesn’t interrupt what they’re watching. In other words, it’s “start this agent over from a blank page.” You’ll get a confirmation prompt, and the action is irreversible.

4. What you can change in the Settings subpage

Here’s everything the agent page → Settings subpage lets you configure:

Setting	What it controls
Avatar	Upload an image or generate one with AI
Name	How the agent appears in your list
Persona	Mission and description — defines the agent’s role and personality
Model tier	`Standard` or `Ultra` (💪). Which AI model backs each tier is decided by UpServe. `Ultra` spends far more SU
Reasoning effort	Visible when the selected model supports reasoning. Choose default (auto), none (fastest, minimal reasoning), low (faster, cheaper), medium (balanced), high (more thorough, higher cost), or xhigh (deepest analysis, highest cost). For none/xhigh, the model automatically falls back to its lowest/highest supported level
Tools	Toggle web search, browser, code execution, etc. on or off. Each tool also has a per-tool “Always auto-approve” button to skip the approval prompt for that tool
Always auto-approve	Permanently delegate approval for specific tools — the agent runs them without asking, including during scheduled or autonomous runs
CU action restrictions	When computer use is enabled, individually allow or block each action type (click, type, screenshot, etc.)
Session Auto-Reset	When to automatically clear conversation history: never / daily at a set hour (UTC) / after idle timeout
Unread Message Limit	If unread messages exceed this count, scheduled, Heartbeat, and webhook auto-runs are skipped until you open the chat (0 = unlimited)
Memory retention period	How long before older memories gradually fade (default 90 days)
Auto Dream interval	Every N responses, automatically consolidate memory. Set to 0 to disable.
Daily background budget	The maximum SU this agent’s background runs (schedules, Heartbeat, etc.) can use per day. Once reached, only background runs pause for 24 hours — messages you send are still handled. Use the account default or set a per-agent limit; 0 means unlimited
Assigned PC Specs	The CPU/memory tier of the computer the agent runs on (see section 8 below)
Tool indicators	Whether to show real-time tool execution indicators (including “Thinking…”) in the chat
Channels	Whether this agent’s results send app push notifications
Heartbeat (autonomous activity)	Set the wakeup interval and active-hours bands
Agent Status	ACTIVE / INACTIVE toggle (see section 1 above)
Skills	Add or remove installed skills
Initialize Agent	Reset conversations, memory, and canvas (danger zone)
Delete	Permanently remove the agent (danger zone)

5. Reviewing runs and cost

On the agent page → Logs subpage (/agents/<agent>/logs) you can see recent usage and SU consumed.

Use the 7D / 30D / 90D tabs to select the time range.

What’s shown:

Total SU used in the selected period
Total requests — number of model calls
By trigger — split across your messages / schedules / webhooks / heartbeat, etc. (mentions and other trigger types fall back to a gray bar)
By cost type — LLM cost vs. infra/tools (sandbox, etc.)
By model — which model was used how much
Daily trend — a small bar chart of recent activity
Schedule run history — when each automatic run happened and how it ended

If your SU are dropping faster than expected, check the “By trigger” breakdown first to see if Heartbeat or schedules are doing more than you intended. A heartbeat tuned too aggressively will keep spending SU even when you’re not around.

6. Tidying automations

Toggling a schedule or webhook OFF stops it from running. To keep the screen clean, delete the ones you definitely don’t need anymore.

Delete a schedule: on the Schedules subpage, hit the Delete button on the card.
Disable webhooks: on the Webhooks subpage, switch OFF.
App push notifications: on the agent page → Settings subpage → Channels section, toggle the app push notification switch off.

For Heartbeat, prefer toggling OFF or clearing the active-hours bands instead of deleting — that way the same profile is ready to go when you turn it back on.

7. Skill Library Cleanup

Install enough skills and things start to pile up — skills you never used, skills that keep failing, two skills that do essentially the same thing. This feature lives on the Skills tab’s “My Library” screen, not in an individual agent or a separate account settings page. Open Skills (/skills) → My Library; a banner at the top shows the number of cleanup candidates, and clicking it opens a modal listing what UpServe scans for periodically, organized as individual cards.

What categories does it detect?

There are six.

Category	What it means
Unused (UNUSED)	Installed more than 30 days ago but never run
Failing (FAILING)	Has a high failure rate over the last 30 days (only judged once there’s enough run history)
Outdated (OUTDATED)	The original skill has been updated but your copy is still on an old version
Deprecated Origin (DEPRECATED_ORIGIN)	The original skill has been marked deprecated and is no longer actively maintained
Duplicate Function (DUPLICATE_FUNCTION)	Overlaps significantly with another skill you already have
Revoked (REVOKED)	Your installed version was confirmed malicious and pulled — flagged immediately (no 30-day wait); update to a safe version or remove it

How to use it

Open the “My Library” tab on the Skills (/skills) page.
At the top there is a periodic-scan toggle. When it is ON, UpServe runs the scan automatically on a regular schedule. When it is OFF, automatic scanning is skipped. When there are no current cleanup candidates, the empty screen shows a “Run Now” button you can use to trigger a manual scan.
When cards appear, select the ones you want to act on using the checkboxes, then click “Apply Selected” or “Dismiss Selected”.
- Apply: carries out the recommended action for each item (remove the installation, pull in an update, etc.).
- Dismiss: skips those items for this round. They may reappear in the next scan cycle.
To clear everything at once, use “Dismiss All”.

To learn how the autonomous skill system improves skill quality behind the scenes, see Skills & the Autonomous Skill System.

8. Adjusting its computer specs (Assigned PC Specs)

Every time an agent runs, it uses a dedicated computer (its working environment). You can pick that computer’s CPU/memory spec from six tiers.

On the agent page → Settings subpage, look for “Assigned PC Specs.”

Tiers	Screen	Best for
NANO / SMALL / MEDIUM	No screen	Lighter, cheaper work (running code, processing documents, etc.)
LARGE / XL / XXL	Desktop screen	Work that needs a screen, like browser automation or computer use, and heavier workloads

The default is “Auto.” Leave it there and the agent gets the default spec UpServe picks.
If you pick one manually, higher tiers cost more SU per minute — each option in the list shows its per-minute SU rate.
You’re not charged while the computer is powered down (asleep). SU are only spent while it’s actually running.
🔴 Picking a “no screen” tier (NANO, SMALL, MEDIUM) takes away that agent’s web browser and screen control. That computer has no screen at all. The settings screen shows the same warning, and moving back to LARGE or above restores them. If the agent’s job involves browsing the web or driving a screen, keep it at LARGE or above.
Agents that must have a screen — such as one whose published web app visitors operate directly — can’t go below LARGE.
XL and XXL are only available if you have a paid subscription or belong to a managed organization.
Changes apply starting from the next activation (restart), not to the computer currently running.

The agent can also propose a spec change on its own

If the agent runs into memory trouble during a task (evidence it needs more), or notices it’s had far more capacity than it’s using for a while, it may propose a spec change in chat. Even then, the agent never changes its own spec — you’ll see an approval request showing the current/new spec and the expected cost change, and it only takes effect once you tap approve yourself.

When a downgrade proposal targets a “no screen” tier, the approval card also lists the tools the agent would lose (web browser and screen control) — so you’re not approving on the SU saving alone.

9. Transferring a personal agent to an organization by drag-and-drop

The fastest way to move a personally-owned agent into an organization. On the Office board, drag a personal employee’s card onto an empty spot in an org section (or onto another standalone employee card already sitting there), and it’s transferred to that organization on the spot.

Drop it directly onto a specific team inside the organization, and two things happen at once — the agent is transferred to the organization first, then immediately joins that team. You only see one confirmation dialog; the two steps run automatically in sequence.

This works on the desktop web Office board (both card view and table view) and in the sidebar. It’s not available in the mobile apps yet.
The only direction supported is personal → organization. You can’t use this to pull an org-owned agent back to personal ownership, or to move it between two organizations — to send an org-owned agent back to a member, see 10. Moving an org-owned agent back to personal ownership below.

Before the transfer happens, a confirmation dialog explains:

This can’t be undone. Getting it back to personal ownership requires a separate transfer by the organization’s owner (OWNER).
Every member of the organization will be able to read the agent’s entire past conversation history.
Ownership moves to the organization — it’s no longer your personal asset.

When you drop it directly onto a team, the transfer can occasionally succeed while joining that team fails. In that case the transfer has already gone through — you’ll see a message, and dragging the agent onto the same team again retries just the team join.

This only appears on accounts that use an organization (for a company or department). To move an entire team — the team and everyone on it — to an organization at once, see the “Teams on an organization account” section of Building Multi-Agent Teams. This section covers transferring one individual employee at a time.

10. Moving an org-owned agent back to personal ownership

You can move an agent that belongs to an organization back to a specific member’s personal ownership. This is available only to the owner (OWNER) of the organization the agent belongs to.

Open the ”⋯” menu on the agent in your agent list.
Click “Transfer ownership (to member).”
Pick the receiving member from the list and confirm.

Before you transfer, keep in mind:

This cannot be undone. Moving it back into the organization requires a separate transfer.
Other members of the organization will no longer be able to see past conversations with this agent.
Credentials and external connections attached to the agent move with it to the selected member’s personal ownership.
If it belonged to an org team, it’s removed from that team.
If any external services were connected through the organization’s account, you’ll need to reconnect them under a personal account after the transfer for the agent to keep using them.
The transfer goes through no matter how many employees the recipient already has, and the transferred employee arrives ready to work — there is no headcount limit.

11. Deleting an agent

On the agent page → Settings subpage → danger zone at the bottom, click “Delete”.

What goes away

All automatic runs (schedules, heartbeat, webhooks, event triggers) stop immediately.
The sandbox (code execution environment) is cleaned up.
If the agent belonged to a team, its team links are removed.
The agent disappears from your agent list.

What stays

Conversation history rows remain in the database (soft delete). There’s no UI to view or recover them yourself; they’re kept to preserve linked records (e.g., who authored which skill). The agent’s Brain (memory) data, however, is permanently and immediately removed on delete and cannot be recovered.
Attachments and avatars stay, so other conversations and shared records aren’t broken.
Skills and canvases this agent created that you’ve published publicly or shared with others remain available.

When delete is blocked

While the agent is running, delete is rejected and the screen tells you why. If the delete dialog closes and you land back on the list, the delete definitely went through — there’s no case where it fails but looks like it succeeded. Wait until the run finishes and try again.
Shared system agents can’t be deleted by regular users.

After deletion, the agent no longer appears in the detail view, logs, or cost statistics pages. If you call delete again on an already-deleted agent, it returns successfully without error.

12. Rating automatic-run results on the spot

When an agent produces something on its own while you’re not watching — from a schedule or a webhook, for example — or right after you approve a risky action mid-conversation, a “Was this result helpful?” rating row appears right below that message (same on web, iOS, and Android).

Tap Useful and it’s saved right away.
Tap Needs fix and it’s saved right away too, opening an optional comment box below where you can note what went wrong. You can send it empty.
After rating, tap “Undo” to go back to unrated.

(Upcoming) Confirming the plan before building

The following isn’t deployed yet as of this update. Once it ships, this section will reflect what actually happens.

When you hand off work that produces something visible — a document, an image, code — you’ll be able to set the agent so it doesn’t dive straight into building. Instead, it lays out a short plan first and asks whether it’s okay to proceed. A new toggle will appear on the agent page → Settings subpage, and it’s especially recommended for agents that face external clients or handle outsourced work.

When it’s on, the agent presents a plan of roughly three to six steps and waits for your go-ahead before starting the actual work. If your request changes partway through, instead of quietly reworking what’s already built, it shows you the updated plan and checks in again. Simple questions and small, quick fixes skip this step entirely.

This setting works by instructing the agent to show its plan first — it isn’t a hard block. In rare cases, the agent may not follow that instruction and start working right away.

Advanced

The rest of this page isn’t required reading. Use it only if you want to look at memory structure directly, tune Heartbeat in detail, or check the fine print on sandbox specs and ownership transfer.

Brain file layout

The Brain panel mirrors the .brain/ folder inside the agent’s sandbox. The notable files:

File	Purpose
`SOUL.md`	Agent identity and working style
`AGENT.md`	Operational rules and behavioral guidelines
`MEMORY.md`	Memory index (capped at 200 lines / 25 KB)
`memory/user_*.md`	Your preferences and context
`memory/feedback_*.md`	Lessons captured automatically after tasks
`memory/project_*.md`	In-progress project info
`memory/reference_*.md`	Source material the agent consults often
`memory/<other>.md`	Any memory file that doesn’t fit the above types

Brain files are private to the agent. To share something with other AI employees, ask the agent to “move this to team-shared memory” — it’ll write to a separate shared store instead of the private Brain.

The panel supports direct file editing. It only wakes the sandbox when you hit refresh; the rest of the time it serves a local cache, so leaving the panel open doesn’t cost anything.

How Heartbeat works

The Heartbeat schedule works like this:

Active-hours bands: a unit of weekdays + start time + end time + interval. You can stack multiple bands.
Interval presets: 30 min / 1 h / 2 h / 3 h / 4 h / 6 h / 8 h / 12 h / 24 h. This is a baseline, not a fixed cadence — it stretches when things are quiet and tightens when there’s activity.
Timezone: defaults to your timezone; editable in the card.
At least one band is always required: the editor enforces a minimum of one active-hours band — removing all bands to disable Heartbeat isn’t supported. To fully disable it, use the toggle on the Schedules subpage instead.

During a heartbeat the agent wakes with a lightweight context (smaller than a normal chat turn) and exits immediately if there’s nothing to do, so empty heartbeats stay cheap. You can verify this in the cost log under “By trigger.”

The six computer-spec tiers in detail

Tier	vCPU	Memory	Screen
NANO	1	1GiB	None
SMALL	2	2GiB	None
MEDIUM	2	4GiB	None
LARGE	4	4GiB	Desktop
XL	4	8GiB	Desktop
XXL	8	8GiB	Desktop

Disk space is fixed regardless of tier — tier only changes vCPU and memory.
When you pick a tier directly in Settings, it’s saved immediately as soon as it passes server-side checks (meets the tool-required minimum, meets paid-plan eligibility) — no separate approval step.
When the agent proposes a tier change on its own, it always requires your approval. A downgrade proposal is rejected if the agent had a memory-shortage incident within the last 7 days (recent evidence of that risk is treated as more reliable than the agent’s current read on things). Either way, the change takes effect starting from the next activation.

What moves (and doesn’t) in a personal → organization transfer

Moves with it: credentials for connected external services, webapp links, its email address, Slack channel/thread connections, and any versions it has pinned in the Skill Library.
Team membership breaks at the same time. Unless you drop it directly onto a team inside the organization, it lands as a standalone employee in that organization right after the transfer.
Doesn’t move: external MCP server connections registered on your personal account (moving these would drag along your other personal agents that share the same server), auto-assignment of external services (OAuth) connected through the organization’s account (an admin has to assign these separately), attribution for past SU usage (the ledger is never rewritten retroactively), and authorship credit for any marketplace listing this agent originally published.
The destination must be an organization you’re already a member of — this action can’t create a new organization for you. System agents, snapshot agents, and agents that already belong to an organization can’t be transferred this way.
No special permission is needed — you’re moving your own personal agent into an organization you already belong to, so regular member-level access is enough. Sending an org-owned agent back to a specific member, on the other hand, is a separate action only that organization’s owner (OWNER) can do.

What’s preserved during a reverse ownership transfer

Organization membership itself doesn’t change — the receiving member stays in that organization after the transfer.
Past usage history (SU-spend records) is never retroactively re-attributed.
Conversation history, canvases, and other content are preserved — other members of the organization just lose visibility into them.

Data retention after delete (what we can confirm)

Agent record: marked status = DELETED with a deleted_at timestamp. Hidden from lists and search right away.
Sandbox (code execution environment): torn down immediately.
Automatic runs: schedules, event triggers, and webhook callbacks are all disabled in one go.
Conversation, canvas, and task rows: preserved. For permanent removal of your content rows, contact support.

The above is what the code (AgentService.delete) actually does today. If that changes, this page is updated alongside it.

The fast-hire onboarding window

A new hire created through fast hire keeps its self-editing profile tools until you confirm (approve the summary it shows you). 🔴 There is no time cap (UPS-1072, removed 2026-08-25) — previously the window auto-closed 72 hours after hiring, but now an unfinished onboarding can sit for days without the tools being revoked or autonomous runs turning on.

The hire-time-plus-72-hours value (rookie_until) is still stamped by the server, but it’s display-only now — once that time passes, the progress card switches to “Stalled” to flag the neglect, and nothing else is gated on it. Only your confirmation revokes the tools; from then on the profile can only be changed the way any other agent’s can — through the Settings subpage.

When the rating row appears

There are two cases where the rating row shows up.

Unattended-run output: results from schedules, webhooks, work reports, or tasks the agent proposed and carried out on its own (initiative). It doesn’t appear for Heartbeat check-ins or agent-to-agent mentions.
Approved actions mid-conversation: if a risky action asks for your approval during a live chat and you approve it, the resulting output also gets a rating row — approval only judges the action beforehand, so it can’t speak to how the result actually turned out. If you decline approval, nothing runs, so there’s nothing to rate.

Ratings are only used in aggregate — the agent doesn’t see its own score or react to it in a later conversation.

(Upcoming) How the plan-confirmation policy works

Technical detail for the “Confirming the plan before building” section above. Not yet deployed to production as of this update.

This setting isn’t a system-enforced approval gate — it’s a single instruction line added to what the agent sees. Whether it actually presents a plan is ultimately up to the agent’s own judgment.
It only applies to normal conversation turns responding to a message you sent directly. It’s not added during unattended runs — schedules, webhooks, or Heartbeat check-ins.
The toggle is per-agent and defaults to off.

Managing Your Team in the Office — see every employee at a glance and act on what needs attention
Building Multi-Agent Teams — group employees into a team so they can collaborate