What Cortex actually does
Cortex is the Cortex section of the dashboard and the engine behind it. Left alone, it runs a set of phases on a schedule and produces things you can act on:
- Learned skills - it notices a procedure the agent has repeated and writes it down as a reusable skill.
- Ideas, brainstorms and prototypes - each one a folder in the agent's workspace with real content, not a one-line note.
- Projects - a promoted item with a README, a phased plan with checkboxes, a status file and a journal.
- Proactive items - grounded suggestions and findings, each carrying a concrete action and a source link.
- Self-improvement proposals - suggested edits to the agent's own operating documents.
- Self-checks - findings about the deployment itself, each with a button to investigate.
- Optimisation passes over the workspace, the skill set and memory.
Those map to the tabs across the top of the section: Pulse, Brainstorming, Ideas, Prototypes, Projects, Learned Skills, Self-Improve and Self-Checks.
The settings table
At the top of the page sits the control surface: a two-column table where every setting is one row. The All agents column is the deployment-wide default, on or off. The second column is the selected agent's override, with three states - Inherit, On, Off. Beside each override, the dashboard shows the resolved value, which is what that agent actually runs with once both tiers are applied.
| Row | What it controls |
|---|---|
| Cortex engine | The master switch. Off for an agent means no cognition at all - no scheduled phases, no manual runs. Off for all agents is the fleet-wide stop. |
| Autonomous | On means the phases run on their cadence unattended. Off leaves the engine available but idle - you can still run a phase by hand. |
| Fully Autonomous | Auto-approves the output that would otherwise wait for a click: learned skills are kept, self-improvement proposals are applied, and actionable items are executed. |
| Cadence | Relaxed, balanced or aggressive. Sets how often each phase runs and the daily budget below. |
An agent set to On runs even when the fleet default is Off. "Cortex off everywhere, on for the one agent I am piloting" is a supported arrangement, and Inherit genuinely follows the default whichever way it points. A footer counts an agent's overrides and offers Reset to defaults to clear them in one go.
Controls respond immediately - the switch moves on click and the request follows a moment later, with a toast if anything fails. The same per-agent master switch is duplicated as a Cortex Engine card on the agent's Overview tab, for when you are already there.
Cadence and the daily budget
Cadence is the single knob that decides how hard the engine works. It sets each phase's frequency and the cap on expensive deep runs per day: relaxed allows 6, balanced 12, and aggressive 24. Aggressive is the default.
A Deep runs today meter shows usage against that cap. When the budget is spent, autonomous phases skip until tomorrow rather than quietly falling back to a cheaper model - the engine would rather do nothing than do something worse. Manual actions still work when the budget is exhausted, but they ask for confirmation first, so an expensive run is never a stray click.
If Cortex feels noisy or costly, move cadence to relaxed before turning anything off. Most of the perceived noise is frequency, not the feature.
Phases and running one now
A collapsible Phases & schedule section lists every phase with its own enable toggle, its cadence, its next and last run with status, a Log button showing that phase's recent runs, and a Run now button. A phase currently executing shows a live spinner and disables its own Run-now button, so you can always tell what is in flight.
Three header buttons cover manual work: Think now runs the whole funnel immediately, and Optimize workspace and Optimize skills run those passes on demand. Both optimisers run in the background - closing the result drawer does not cancel the run, and clicking the button again reopens the live status rather than starting a second run. Activity opens the audit log of everything Cortex has done on its own.
Working with what it produces
The Brainstorming, Ideas, Prototypes and Projects tabs are lists of full-width cards, each showing a description, when it was last touched, its source and its tags. Every section has Active / Promoted / Discarded sub-filters with counts.
Click a card to open it. From there, or from the card itself:
- Open - the detail plus the files in the item's folder, and a project's progress bar driven by its plan checkboxes.
- Explore more - a deeper research pass that enriches the folder.
- Create Project - promotes the item and has the agent scaffold a plan around it.
- Chat - opens a conversation seeded with the item's context, which is usually the fastest way to interrogate an idea.
- Discard - removes it from the list. It stays on disk.
- Send to Agent - on prototypes, asks the agent to run or iterate it.
Your place is preserved: the active tab, the sub-filter, an open drawer including its in-drawer conversation and unsent draft, and your scroll position all survive switching to another section and back. Acting on a card refreshes the list in place rather than jumping you to the top.
You do not have to use the dashboard at all. Agents with the engine enabled carry the matching chat tools, so "show my ideas", "promote the pricing idea into a project", "explore that brainstorm further" and "discard the second prototype" all work in conversation and operate on the same items the dashboard shows.
Approving, undoing, and Self-Checks
Every autonomous change is snapshotted with a one-click Undo. Changes to an agent's identity documents are treated differently again: they always land as proposals in the approval queue - applied automatically when Fully Autonomous is on, and held for your click when it is off. Turning the setting on sweeps whatever is already pending.
Self-Checks findings are actionable rather than informational. Each one carries Fix with agent, which opens a conversation seeded to investigate and, where safe, repair the issue, and Dismiss for findings you have judged and rejected.
Steering it
Below the settings panel, a collapsible Customize your Cortex Engine section edits the agent's creative-direction file without leaving the page. It has three free-form parts - Goals, Dreams and Avoid - and steers the generative phases only. It does not change how the agent behaves in ordinary conversation; that is what the system prompt and identity documents are for.
The file renders as a markdown preview by default with an Edit button that swaps in a raw editor. A badge marks it as customised once it holds real content. This is the highest-value ten minutes you can spend on Cortex: an engine told what you are actually trying to achieve produces relevant ideas, and one left on the default scaffold produces generic ones.
A sensible starting configuration
- Enable Cortex for one agent first. Leave the fleet default alone and set that agent's override to On.
- Leave Fully Autonomous off to begin with. You will see the proposals in the queue and can judge the quality before letting them apply themselves.
- Set cadence to relaxed. You are evaluating output quality, not throughput.
- Write the creative-direction file. Goals, dreams, and the things you never want it working on.
- Come back in a week. Read Activity and the Ideas tab. If the output is useful, raise the cadence or turn on Fully Autonomous; if it is not, sharpen the direction file rather than turning the engine off.