A telco's learning layer for agentic development

Hyperstruck is the intelligence layer under your agents. Learnings are what to do next time. Facts are what is true about a customer, invoice, or system. A telco has partnered with us to be that layer for agentic software development. Their teams can ship more, with the controls they need.

A telco made Hyperstruck their learning layer

A telco has partnered with Hyperstruck as the learning layer under their agentic software development. Agents keep lessons and facts across people and tools, so output compounds and the work stays inside the rules they set.

Example: an engineer teaches an agent a repo rule on Monday. The next engineer, on a different machine, starts with that rule. They do not spend the morning re-discovering it.

If you want the same layer under your own agentic work, write hello@hyperstruck.com.

Ask about an account. Get the facts from its file.

You can ask a question and get the facts Hyperstruck holds that answer it. Those facts can come from that account's own documents and from the meeting notes that concern it. One ask can brief the agent on both.

Example: "What did we agree with Acme?" can pull payment terms from the MSA and the last QBR notes, not only from one or the other.

Agents can now list pending obligations

Each agent's tab shows its own open count. You can see, search, and clear those items in the dashboard. Commitments the agent itself spoke, and ones harvested from a file you uploaded, can land there. A due date keeps the stated day and reads in the writer's timezone.

Example: "send the quote before Friday" is due before Friday for the sales rep who wrote it. Their lead can see three open items on that agent and clear a stale one.

Agents can find what was left unsaid

An agent can name what it found and what was not said. You can see the gap without re-running the search.

Example: a legal agent asked about termination and late fees can say it found the termination clause and did not find late fees.

A fact cites the sentence, not the filename

You can submit a document and point a learning at the passage it came from. Your own code can do this through the Documents API. We wrote this up.

Example: a collections rule points at the net-60 sentence in the signed policy, not at a file named policy.pdf.

Delete the file, and those citations go with it

A document keeps one identity across uploads. When you erase it, the receipt counts what was reached. Delete an agent and its documents go with it. Rotating a key is not treated as a delete.

Example: legal withdraws last year's MSA. Facts that pointed at it show as withdrawn. The new MSA is a new source.

Recalled facts can name their source

A recalled fact can name the meeting or document it came from. You can record a meeting, its attendees, and a due date set as "before a weekday." A meeting can hold several attendees, not only the last name written down.

Example: a QBR with Jane, Raj, and the CFO stays as three people. A later collections question can say the payment-terms fact came from that meeting.

New facts are used until overridden

When two values sit on the same field, the newer one is used until someone overrides it. A fact that just landed can be used right away.

Example: last month's note says Customer X is on net-30. This week's ERP export says net-60. The agent uses net-60 until finance overrides the field.

A corrected fact does not re-teach the old rule

A corrected fact no longer feeds the rule it once taught. A rule also drops credit that came from evidence you erased. A curator can settle which stated rules still stand.

Example: finance corrects Customer X from net-30 to net-60. The old chase-on-net-30 rule does not keep teaching itself to the next collections agent.

A failed send can hold the next run

A rule learned from a tool failure can now hold the next run until the missing field is there.

Example: a payouts send fails because the sort code was not read. The next run can wait until that field is present, instead of sending again.

A team can have its own space

A key can create, own, and delete its own space. A large tenant can still find the right agents. An owner with no memberships can invite themselves in.

Example: collections and support keep separate spaces. A platform lead can still list both.

You can see the learning quota you have used

The stored learning quota is measured, so you can see usage instead of guessing.

Example: a platform lead can see the collections agent is near its cap before a big policy ingest.

A document can belong to an account

A connector can say which account a document belongs to. A question about one account then reaches that account's own documents.

Example: the Acme MSA is marked as Acme's. A question about Brightside does not pull Acme's terms.

A policy upload is ready in about a second

An uploaded document is ingested, not only accepted. Ingest now starts in about a second, not a couple of minutes.

Talk to us

Want the same learning layer that telco is running? Write hello@hyperstruck.com.