Flows: automate work across apps

What a flow is

Chat is you and a model, turn by turn. A flow is the same capability with nobody at the keyboard: you decide once what should happen, and Akili repeats it every time the trigger fires. A weekly digest that reads a knowledge base and emails a summary; a lead-enrichment flow that wakes up when a form posts to a webhook, researches the company and drops a note in Slack; an incident summary an agent can start from chat. Each run is recorded step by step with its inputs, outputs and credits, so you can see exactly what happened.

Flows appears in the sidebar when your plan includes it. Operators of a self-hosted tenant build turn the module on with FLOWS_ENABLED=true; the SaaS build has it on. Where the module is off, the Flows page says so and nothing else changes.

Create a flow

Three ways in, all from the Flows home:

1. Step 101

   From scratch

   Press New flow. You land in the builder with an empty canvas; add a trigger first, then the steps.
2. Step 202

   From a template

   The templates row on the home page offers 7 starting points: Weekly digest, Lead enrichment, Incident summary, Document ingestion, Auto-repost, Auto-plug, RSS to drafts. Each card lists what the template needs before it can run, in the same wording the builder uses. Use template creates a draft copy under your name.
3. Step 303

   Draft with AI

   Describe the outcome in plain words. The assistant may ask a question or two with quick replies, then proposes a chain of steps and a short Before this can run list. Nothing is saved until you press Insert; if the proposal used a node your plan lacks or more nodes than your plan allows, a notice says what was dropped or trimmed. Draft with AI is a plan feature on every paid plan.

The builder

The builder has four tabs: Build, Runs, Triggers and Settings. Build is the canvas.

Palette, canvas, inspector

• Palette — a searchable list grouped as Trigger, Agent, Action, Logic and Apps & connectors. Click, press Enter or drag to add a step. Each entry shows whether it is ready for you, needs a connection, is not in your plan, is not granted to you, or is not available in this deployment.
• Canvas — drag from a node's handle to another node to draw an edge. Edges into a trigger, self-loops and duplicates are refused as you draw; a cycle is refused when the draft saves. A condition offers yes and no handles, a loop each and done, an approval one per outcome, and every node a failure handle for an error branch. Press L for a list view of the same flow when the graph gets wide.
• Inspector — select a node and configure it on the right. The inspector has Configure, Test and Output tabs, and shows readiness problems for that node as you type.
• Notes — sticky notes on the canvas for the next person who opens the flow.

Saving

The draft autosaves 1.5 seconds after your last change; ⌘S saves at once. A save line under the title shows saved, saving, failed or a conflict when someone else edited the same draft, and the builder shows who else has it open. Offline, edits stay on your device until you reconnect.

Draft, published and readiness

A flow is draft until its first publish, then active, paused or archived. Publishing freezes the draft as a numbered version with a one-line note, activates its triggers, and opens a fresh draft on top. Triggers always fire the published version; the draft is yours to break.

Before it publishes, Akili checks the draft and lists anything that would make a run fail. Blocking problems stop the publish with the sentence to fix; advisories publish with notes. The checks you will meet most:

• Plan limits — more nodes than your plan allows, or a schedule tighter than its shortest interval.
• Graph — a cycle, an invalid definition, a step no trigger reaches, or an approval outcome with no edge.
• Access — an agent you are not granted, a knowledge base that is not yours, a skill with no approved version, a tool that is unknown or turned off for the workspace.
• Connections and identity — a missing MCP connection, an app action whose connection does not exist or lacks a scope, or a step set to run as an identity that has no credential for an unattended trigger.
• Triggers — a webhook with the raw body format, a webhook on a plan without webhook triggers, an event the catalog does not offer, or an agent tool name that is reserved or already taken.
• Advisories — an agent step still carrying the palette's default instructions, an approval with no deadline (it will expire after 30 days), a literal secret in an HTTP header, or a search tool served only by your personal credential, which means colleagues cannot run the flow.

Pause stops the triggers without unpublishing; Activate resumes and resets the failure counter. A flow also pauses itself after three consecutive unattended failures (configurable from 0 to 20 under Settings), when a run fails for lack of credits, or when an agent credential it depends on expires; the Flows home shows why. There is no unpublish: restore an earlier version into the draft from Version history and publish it, or archive the flow.

Test before you publish

Select a step and open the inspector's Test tab. Give it a trigger payload as JSON, or pick a sample: the last captured webhook request, a simulated schedule firing, the catalog's sample for an event, or the inputs an agent would pass. Run test executes that one draft step with the real executor and shows the result, input and output. Upstream values come from pins (an output you pinned as test data from a run or a previous test) or from each step's latest test within the last day; the tab names the steps you still need to test first. Step through tests the selected step and everything after it, stopping at the first failure.

A step test is not a run: it never counts against your runs per day or concurrent runs, and test rows are cleared after seven days. It does spend the credits the step spends. Only the owner or a flows admin can test steps. A full run of a draft is not possible; publish, then press Run (or Run with input… when the flow defines inputs) and watch the run page.

Limits

The Max plan sits between Pro and Team (10 flows, 50 nodes, 30-minute schedules, 200 runs a day, 60 days of history). Free plans do not include Flows. Beyond the plan table: a run times out after an hour by default (settable up to 24 hours; time spent waiting or on an approval does not count), a run executes at most 1,000 steps, a step's output may not exceed 1 MB, a wait may not exceed 30 days, and pinned test data stays under 256 KB. A run past the daily limit is refused with a rate-limit error, and a run past the concurrency limit is refused until one finishes.

Sharing and permissions

Everyone in the organization who can read flows sees every flow. What they can do with it comes from their role: members can create, edit and run; analysts read only; admins and owners also archive, publish, pause, delete and share any flow. Share on a flow gives one member more than their role gives, never less: Can run adds start, cancel, replay and firing triggers; Can edit adds everything but archiving, deleting and sharing. Publishing, pausing and testing steps stay with the owner and flows admins.

Good to know

• Credits. Each step records what it spent; the run row shows the total, or spent versus estimated while it runs. Set a cap per run in Settings or per agent step in the inspector; the estimate panel tells you when the estimate exceeds the cap.
• Several triggers. One flow may have a schedule and a webhook and an agent trigger. The Triggers tab shows each one's state, next firing, last run and how many requests it rejected.
• The owner's personal connections. Steps that run as the owner use your Google, Slack or Teams connection, which is only allowed on manual runs. Unattended flows need an agent with its own credential; see Runs-as.
• Mobile. The Akili app lists your flows and runs and lets you decide approvals; building stays on the web.
• API. Keys with the flows:read, flows:run and flows:write scopes list, start and manage flows; see Runs and approvals.