AGENTS & AUTHORITY

Agents that do real work, within strict limits

Create an agent by describing the job, then decide what it may create, read, update and delete. An agent can never have more access than the one above it, and everything it does is recorded.

Daya
Agents Chats Emails Calendars Recordings Integrate
WORKSPACE
Base apps6
Mini apps5
Drive21
Analytics23
ATTENTION
To-dos inbox15
@ / Unread11
Flagged2
Drafts3
STREAMS
payments-revamp
launch-prep
announcements
watercooler
platform
DIRECT MESSAGES
System agent
Maya
Ravi
Support agent
Research agent
Finance agent● Active
Answers to Mia Tan · in 3 streams
PauseEdit authority
AuthorityCognitionActivityMemory
WHEREREADDRAFTACTDELETE
Refund trackerapproval—
Stripe refundsapproval—
Ledger routing—approval—
finance@acmeapproval—
# payments-revamp—
QuickBooks———
Anything over $500 or outside these rows goes to a person first.
Spend cap$5,000 / week
$1,904.20 used · resets Monday
Waiting for approval
Release refund R-1047
$980.55 · Selin Kaya · asked Mia 4m ago
ApproveDecline
CognitionClaude
FastCareful
REFUND TRACKER
CREATE
READ
UPDATE
DELETE
Humans
✓
✓
✓
✓
System agent
✓
✓
✓
—
Finance agent
✓
✓
✓
—
Research agent
—
✓
—
—
Set for each person, role and agent, and for each base, channel and app.

Access control as precise as it needs to be

For every person and every agent, set what they may create, read, update and delete in each part of the workspace. Anything not allowed is out of reach.

Set per person, role, agent, base, channel and app
An agent can only pass on access it has, so a sub-agent never gets more than its parent
New agents stay in draft until a person approves them
A change to the hierarchy has to be approved by a person

An org chart of agents with people at the top

Describe an agent and it appears on the org chart as a draft. The system agent coordinates the other agents, which each run workflows within their access and spending limits. Every handoff between them is logged.

A spending limit for each agent, with approval needed above it
Skills, tools, policies and knowledge set for each agent
Escalation always ends with a person
Agent hierarchy1 change proposed
Kate Sullivan owner
System agent workspace
Support agent
capped at $200/mo
Finance agent
approvals on
Research agent
read only
Every handoff logged · escalation ends at Kate

What controls an agent

The matrix

What each person and agent may create, read, update and delete, in each part of the workspace.

Spend caps

A spending limit for each agent, and approval for anything above it.

Policies

Company and agent documents that an agent has to follow.

Knowledge base

What it reads before it drafts anything for someone outside the company.

Skills and tools

Upload a skill or give it a tool, and take either away later.

Audit trail

Every action, handoff and approval is recorded once and cannot be edited.

Finance agentthinking · 14sStop
Kate: which refunds have been held more than five days?
✓ Read refund tracker · 28 rows this week
✓ Found 3 held more than 5 days · $3,620.55
✓ Checked all 3 against the processor
◌ Drafting the note to finance
Anyone who can post in the stream can stop an agent working there.

Watch an agent as it works

While an agent is working, it shows what it is doing now, how long it has been going and what each finished step found. Anyone in the stream can stop it, and the steps it finished stay on the record.

A live list of steps while it works, not a loading spinner
Stop it part-way and the record of what it did remains
Try again starts from the question, not from scratch
Each run shows its cost, step by step
COGNITION SETTINGS

Choose the AI model for each kind of task

Permissions decide what an agent may do. The model chosen for it decides how well it does it, and what it costs.

ClaudeGPTGeminiDeepSeekKimiGLM
Routing and taggingAffordable
Customer repliesValue
Reasoning and orchestrationExpensive
Document and image readingValue
Code and scriptsExpensive
The same model can be good value for one task and expensive for another, so each model is rated for each task.

Eighteen kinds of task, and a model for each

There is no need to use one model for everything. Daya splits an agent's thinking into named tasks, such as routing, retrieval, perception, generation, reasoning and governance, and offers a short list of suitable models for each.

A default for the workspace, which a single agent can override for a single task
Each option is labelled Affordable, Value or Expensive for that task
Change a model at any time without rebuilding the agent
Prices and limits are listed on every model's page

Pay for AI work in advance, with a fixed ceiling

The subscription pays for seats. Agent work is charged by usage, from credit bought in advance, and the balance shows how many days it will last at the current rate of spending.

Charged at each model's published price, with nothing added by us
Automatic top-up has a monthly ceiling and a waiting period between charges
Costs are recorded for each workflow step, so a spike can be traced to its cause
AI USAGE CREDIT
$4,280balance
About 26 days at the current rate
Auto top-up$500 at $1,000
Monthly ceiling$3,000
Cooldown between charges24 hours
At the ceilingNotify, do not charge
At zero, agent work stops and nothing is billed later, so the balance is shown in days as well as dollars.
IN THE BOX
Create an agent by describing it
Per-surface authority matrix
Strict-subset delegation
Draft until approved
Spend caps per agent
Approval workflows
Agent hierarchies
Policies and knowledge base
Skills and tools per agent
Design system per agent
Full audit log
Escalation to a person
18 cognition roles
Multiple model providers
Bring your own keys
Per-agent model overrides
Prepaid AI credit with a ceiling

Common questions

Do agents take a seat?

No. Agents are part of the plan. Seats are for the people who log in.

What stops an agent going too far?

Three things: its permissions limit what it can touch, its spending limit caps what it can cost, and anything that cannot be undone needs approval first.

Can an agent create another agent?

It can propose one. The new agent cannot have more access than the agent proposing it, and a person has to approve it before it runs.

WORKS WITH

Give an agent a job and its limits together

Start with one agent, one base and read-only access. Give it more once you trust it.