$18M a year of find-and-figure-out work

Hyperstruck is the intelligence layer under your agents. When a run finishes, two things stay: lessons (what to do next time) and facts (what is true about a customer, invoice, or system). A 100-person organisation where people manage 3 agents each is looking at about $18M a year of find-and-figure-out work. This cycle you can put that number on a board, those facts follow the agent into any editor, and agents finish work nobody could script in advance.

$18M a year, on a board you can show

The dashboard now offers applied-lesson and time-saved charts. Time saved is counted once per successful run. Recall can be attributed to the editor that made it.

A 100-person organisation where people manage 3 agents each is looking at about $18M a year of that work. A 500-person organisation is looking at about $90M. You can put that number in front of a steering group.

Facts follow the agent into any editor

Agents in Cursor, Claude Code, and Claude CoWork now receive the facts they already established, and earn credit for using them. You do not re-teach the same customer record because someone switched tools.

Example: a collections agent already knows Customer X is on net-60. The next person who opens that work in a different editor starts with that fact. They do not chase on net-30, and they do not spend the morning re-checking the ERP.

Dangerous facts are treated as dangerous

Hyperstruck now distinguishes remembered data that is costly to get wrong. A wrong bank sort code is protected with the same seriousness as a wrong US routing number. Tool authors can mark a field as harmful if the value is wrong.

Example: a payouts agent writing a UK sort code or an Australian BSB does not treat that field like a display name. A wrong value here sends money to the wrong account.

Agents finish work nobody could script

Agents can complete work whose next step depends on what just came back, not on a path written in advance. They stop repeating a call that never returned. If a plan fails, the retry uses what failed, not only the original ask. If the agent planned a fallback, it takes that fallback and finishes.

Example: an investigations agent following a chargeback keeps going from the bank's actual reply. If the CRM lookup timed out, the next attempt accounts for that. If the primary billing API is down, the agent uses the backup path it already planned, instead of looping until someone kills the run.

Credit only when the lesson reached the agent

A lesson earns credit only when the agent actually saw it. A run can name the stored facts it used, and can say when it used none.

Example: "always check the bank statement before chasing" does not count as applied unless it reached the agent on that run. A steering group sees real use, not a full shelf.

Customer facts stay with the right agent

Shared lessons do not carry another agent's customer data. Another team cannot read, change, or delete your agent's facts by pointing at them. A one-off exception also stays an exception until it has earned the right to widen.

Example: finance's payment terms for Customer X stay in finance. Support does not inherit that record. "Do not chase Customer X this week" does not become "do not chase anyone on net-60."

Later lessons still get used

A large lesson library still gets used in full. The tenth collections rule still has a chance to fire, not only the first few.

You can operate the learning loop

You are told within a day when learning credit stops being earned. Each uncredited run names its cause. One-shot runs and the last turn of a session report what they learned. You can search the lesson inventory in plain text, and open a Needs review list.

Example: a platform lead searches "net-30" and sends stale payment-term lessons to review before the next collections cycle.

Space keys, and a tighter security bar

Team leads can manage agents in the spaces they own, without an org-wide write key. Hyperstruck also tightened the security bar: secret scanning a pull request cannot exempt itself from, deploys that refuse a lent environment, and tool connections that use the tenant's own credential.

Example: the finance space manager can create and update finance agents. They cannot rewrite legal's fleet.

We measure whether a lesson transfers

We measure whether a lesson still holds on work the agent has never seen. That is the number a procurement conversation can use, because it is about new work, not a replay of the training set.

Talk to us

Want to see this on your own workflow? Write hello@hyperstruck.com.