By Yuvan Sundrani · August 2026 · 5 min read

Why we built this

When you run a flow on Autosana, the agent breaks down your natural language instructions, reads the screen, and reasons through each step using inference. That reasoning is what makes flows resilient: the agent understands intent, adapts to UI changes, and navigates the app the way a real user would.

But when a flow passed yesterday and the app has not changed since, there is no reason to reason through every step again. The agent already knows exactly what to do. Running the full inference pipeline on unchanged screens is redundant work that costs time and money.

Run Caching eliminates that redundancy.

How it works

After a flow passes successfully, Autosana stores the exact action sequence the agent executed. On subsequent runs, the flow replays that cached sequence instead of running full inference on every step.

This is where it gets interesting. When something in your app does change, the cache does not just break and start over. If replay cannot continue — for example because the UI looks different — the full vision-based agent takes over and reasons through the flow normally.

Caching is best-effort and enabled by default for flows created in the dashboard. After a flow's next successful run, subsequent executions will automatically use the cache. Flows created via the API default to caching off and can be enabled from the dashboard. The org-wide toggle lives under Settings → Agent Features.

What this changes for your team

A regression suite that previously took an hour to complete now finishes in minutes, because the majority of flows in a nightly suite are running against an app that has not changed since the last passing run. Those flows no longer need full inference on every step, and the time savings compound across the entire suite.

Pull request smoke tests return results faster. Engineers waiting on test results before merging get feedback sooner, which means less idle time between writing code and shipping it.

Cost also drops meaningfully as suites grow. A 100-flow regression suite running every night no longer carries the inference cost of 100 fresh runs. Flows that have not changed since their last successful execution run at a fraction of the cost, and the savings scale with the size of your suite.

None of this comes at the expense of correctness. Cached runs replay the exact sequence that passed previously, and if the UI looks different the full agent takes over automatically.

A few things worth knowing

Setup and teardown hooks do not affect caching. Pre-test data creation and post-test cleanup continue to work exactly as before.

Runtime hooks and Set Variable actions require full inference. If a flow uses a hook as a mid-flow step or sets a variable during execution, that flow runs with the full agent because its behavior depends on values determined at runtime.

Dynamic values from setup hooks need to be exported. If a setup hook generates a value that changes every run, such as a test email or authentication token, write it to /tmp/autosana.env so Autosana can track it across runs. Values that are only printed without being exported make every run look unique to the cache, which prevents it from engaging.

When those exported values change every run, the cache takes three passing runs to settle. The first records the action sequence, the second detects the changing values, and the third uses the cache. Failed runs do not count. Editing a flow's instructions resets the count. A flow without changing hook exports uses the cache on the run after its next successful run.

Get started

Run Caching is live and deployed for all customers. Your flows will begin caching automatically after their next successful run.

Full documentation: Run Caching and Hooks and Run Caching.

Get started at autosana.ai.

Prefer a walkthrough first? Book a demo.