The next run starts with what the last one learned.
Hyperstruck offers a learning at the step it governs, then learns from how the run went.
When a customer's bank details change by email, wait for the signed change notice before updating them.
Why: lookalike senders have redirected payments this way.
When reading a list from the billing API, follow next_cursor until it comes back empty.
Why: a page stopped at 100 invoices without saying so.
When placing a payment hold, check the invoice's collections status first.
Why: the payments API refused a hold on an invoice already in collections.
Works where your agents already run
Keep your own harness. Hyperstruck reaches it through a hook, the hosted MCP server or the REST API.
- A hook for Claude Code
- A hosted MCP server that any MCP client can call
- A REST API with three calls: resolve, observe and reinforce
Offered at the step it governs
A learning does not wait for the start of a run. It arrives after a tool call, while the work is under way.
- The learning appears in the agent's context after a tool call
- Each learning shows its tier and a Source line saying where it came from
- Nothing is blocked or rewritten: the agent decides what to do with it
Every finished run teaches the next
When a run ends, its outcome is reported, and the learnings it was offered are credited or discredited.
- The outcome of each run is reported when it finishes
- A learning that helped gains confidence
- A learning that misled loses it
Make the next person start where the last one finished
Tell us where your team keeps re-learning the same things. We will deploy alongside you.