Accounting Flow

Saved steps: an agent library your firm can reuse

· 5 min read

AI agentsAutomation

The instruction that took three attempts to get right should not live inside whichever agent happens to have it. Save a step once, drop it into any flow — and it copies rather than links, on purpose.

Firms were writing the same wording into every agent that needed it. The version that took three attempts to get right ended up inside whichever agent happened to have it, where nothing else could find it, and the next person wrote their own slightly worse one.

A step is an instruction with a name

A saved step is one instruction, named and described, living on Agents → Steps. Drag it onto any agent's canvas and wire it into the flow.

What a step deliberately does not carry is a schedule, a client filter or email settings. Those belong to the agent using it. The same "is anything overdue?" step is worth running weekly against VAT clients and monthly against payroll ones — and a step that carried its own targeting could only ever have been one of those.

Copied, not linked — on purpose

Dropping a saved step into a flow copies its wording. The agent owns that text from then on and can edit it freely, and editing the library later does not reach back into agents that are already running.

That is the deliberate choice, and it is worth saying why. A library step is a starting point somebody tidies up months later. An agent quietly changing what it does — on a schedule, unattended, against the whole client list — because a colleague reworded a shared step is exactly the surprise this feature must not produce.

Deleting a step from the library cannot break a running agent. The library shows a usage count rather than pretending to be a dependency.

The flow is the agent

A flow is stored on the agent, so everything it already has comes along unchanged: which clients or jobs it runs for, its schedule, its safety limit per run, where results land, who gets notified. Only the shape of the work changes.

An agent with no flow is the straight line it always was. The two are alternatives rather than a migration, and clearing the canvas gives you the original agent back rather than a lossy translation of it.

And when it runs

Agent schedules follow the same repeat logic as workflow tasks, including the two patterns you cannot express as a day number: the last working day of the month, which steps back off a weekend, and the nth weekday — "the second Tuesday", "the last Friday". Narrow any month-shaped cadence to chosen months and you have quarterly and annual without needing separate frequencies for them.

See what your firm could save

Two minutes, and it uses your own numbers.

Open the ROI calculator