Identity
Open Agents and click a card to reach the agent's detail page. The Overview tab is the editable surface, and every field on it saves as you go - selects and toggles on change, text fields when you click away. There is no separate edit dialog to remember to submit.
The Avatar card takes an uploaded square photo, falls back to an emoji, or can Generate cartoon - a portrait drawn from an optional reference photo, a description, a style note and an emoji vibe. The portrait comes back within a minute or so. The Identity card holds the model, the description, the accent colour used to tint the agent everywhere in the dashboard, and Access: whether the agent's environment is read-only, read and write, or read, write and execute. Start read-only for anything customer-facing, and widen it only when a task genuinely needs to write files.
The description is not decoration. It is what other people - and other agents - see when they are choosing who to ask, and it is what the agent-picker search matches on.
Standing instructions
The System prompt card is the agent's brief: who it is, what it is responsible for, what it must never do, and how it should answer. It is the single highest-leverage field on the page. A Generate with AI button drafts one from the agent's name, description and selected tools; treat the result as a first draft to edit rather than a finished brief.
Good instructions are concrete and bounded. Name the audience, name the tone, name the escalation rule ("anything about refunds, hand to a human"), and name the things that are out of scope. Vague briefs produce agents that are confidently wrong about the edges of their job.
Editing the system prompt does not need a restart. Running conversations get a one-off note that their operating instructions changed and pick up the new text on their next turn.
Choosing a model
Olano runs across 18+ LLM providers, and the model is a per-agent choice. On a managed deployment the dropdown leads with the Olano-managed tiers in a fixed ladder - Olano Turbo (fast and inexpensive, the default), Olano Pro (harder multi-step work) and Olano Max (maximum capability, highest cost) - each billed in credits with no provider key of your own. Below that sit any providers you have authenticated yourself.
The dropdown only lists providers you can actually run: a key-based provider appears once its key is present, OAuth providers once connected. You will never pick a model that cannot answer. Where a model supports reasoning-effort control, a Reasoning effort dropdown appears under Model; models without it show nothing, because there is nothing to set.
You do not have to commit. In any conversation, the model chip in the chat header runs that one conversation on a different model, leaving the agent's configured default alone.
Tools
The Tools tab is a searchable browser over everything this agent could be given: 75+ built-in integrations exposing 450+ connector tools in total, grouped into categories and curated bundles. A left rail filters to All, Selected or Missing keys; the main pane lists checkbox cards with a readiness badge - Ready, or a count of the credentials still missing - and names the secrets each tool needs. Selection saves itself.
Some tools are injected automatically and are not listed here: memory, time, web lookup and a handful of other universal helpers that every agent gets. Operations tools - scheduling, approvals, heartbeat management - live on a built-in operations subagent rather than being handed to every agent directly.
At the bottom of the tab, a collapsed Approval overrides (advanced) section lists the selected tools with their effective approval state and a per-tool Default / Always ask / Never ask control. That per-tool setting is the highest-precedence rule in the approval system - see chapter 8.
More tools is not better. Every tool the agent holds is a tool it can reach for at the wrong moment, and a longer list makes each individual choice less reliable. Give an agent the tools its job needs and nothing else.
Skills and subagents
A skill is a written procedure the agent can follow: a folder with instructions and, optionally, scripts. The Skills tab lists the skill directories available to this agent and lets you enable or disable them. Skills are how you teach a repeatable method - "how we qualify an inbound lead", "how we format the weekly report" - without stuffing it into the system prompt. They are covered in chapter 3.
The Subagents tab holds nested specialists this agent can delegate to. Each has its own instructions, its own tools and optionally its own model, and runs in its own context so the parent's conversation stays clean. Add one from scratch, from an expert archetype, or as an integration tied to a single service. Delegation is covered in chapter 4.
Per-agent settings
The rest of the Overview tab is a stack of option cards. Every control carries a hint saying what it does, and every dropdown lists what each option means - you should not have to guess.
| Card | What it controls |
|---|---|
| Memory | Whether the agent keeps long-term memory and searches its own documents. See chapter 3. |
| Heartbeat | Turns on recurring background checks and picks how often they run. Applies live, no restart. |
| Oversight | The persistent approval baseline: a mode of Trusted, Supervised or Locked-down, plus "ask before write actions" and "ask before shell commands" toggles that layer on top. |
| Cortex Engine | A master switch for this agent's autonomous improvement work. Off pauses it immediately. |
| Multi-user mode | Turns one agent into one many people can share, with memory isolated per person. See chapter 9. |
| Access & Safety | Human-only (no other agent may message it), Hidden (withheld from listings), auto-accept in chat, and numeric limits on how long a single turn may work. |
| Rate limits | Caps on inbound messages per minute, hour and day, with optional per-channel overrides. Scheduled and autonomous work is exempt. |
| Schedule | A full editor for this agent's recurring jobs. Saves apply live. See chapter 6. |
Further tabs cover the rest: Connectors and MCP for external reach (chapter 5), Webhooks for inbound and outbound events, Commands for which slash commands chat users may run, Talk to Agent for org placement and messaging permissions, Variables for credentials scoped to this one agent, and Advanced for every remaining setting. Where your deployment allows it, a Config tab shows the agent's underlying configuration file directly.
What makes a good agent
- One job, stated in one sentence. If you cannot describe the agent's remit in a sentence, split it into two agents.
- Instructions that say what not to do. Scope limits and escalation rules prevent more bad outcomes than capability lists create good ones.
- The smallest toolset that works. Add tools when a task fails for want of one, not in anticipation.
- An oversight setting you would defend. Read-only research can run loose. Anything that sends, posts, pays or deletes should wait for a human until you have watched it work.
- A model matched to the task. Everyday triage does not need the most expensive tier; a weekly analysis might.
Agents can be stopped and started from the top bar of their detail page. A stopped agent stays stopped across restarts and shows a paused badge until you explicitly start it again - so pausing something noisy is safe and durable.