Real work runs through stages, each with its own procedures, tools, and objectives. Now your AI can too.
Most AI agents are a loop: a model, a prompt and every tool, repeated until the task is done. For work that repeats, that means a new plan on every run, and nobody reviews it first. Malleable starts from the same loop and changes seven things. What you end up with is a harness.
The Malleable team · 3 minute read
Work moves one stage at a time.
The AI is in one stage at a time. It can't leave until the stage has the data it needs, and it can only go to a stage the plan connects.
How a run works →Each stage gets only its own tools.
The stage that reads an invoice can't post one. If the PDF hides an instruction to pay it now, there's no tool for that.
You decide what can't change.
Leave a tool open, lock some of its inputs, like the account to post to, or turn it into a script that runs the same way every time.
What the AI can and can't do →Decisions go to the right person.
A decision that belongs to a person becomes a Handoff: a form made for that decision, sent to the right person in Slack, Teams or email. Anyone can answer, no account needed.
How Handoffs work →Runs can wait for days.
While it waits for an answer, it sends reminders, escalates on a timer, and picks up where it stopped, even through our software updates. Waiting uses no computing power.
Separate agents plan, run and review.
The architect builds the plan with your team. The runner follows it. The reviewer reads real runs and suggests fixes. A run follows a plan instead of making one up.
A person approves every change.
Every change, even the reviewer's, starts as a draft. It gets a test run, and a person publishes it. Runs in progress stay on the version they started on.
Drafts and version history →A custom harness for each workflow.
Same model, different software around it. Your team reads it on the Canvas, and it's the thing that runs. Better models make each run smarter. They don't add stages, Handoffs or drafts.
The ideas behind it.
None of these ideas are new. Researchers worked them out decades ago by watching programmers, office workers, surgeons and spreadsheet users. What's new is AI that can follow them.
See what we kept seeing, and the research behind it →Who built it.
We're a small team that has built workflow software together, with a combined three decades at Asana. Eric Pelz led Asana's workflow group. The teams he worked with weren't asking for a platform to build a solution. They wanted a solution to their problem. So we built one.
Our office is in downtown San Francisco, near Powell station.
More about us →
See your process in stages.
Describe one workflow. Read the plan before anything runs.