OSAgentOSAgent

Ask the record

Ask your company. Get the record.

The answer is already in your workspace — spread over a thread from Tuesday, a pull request nobody linked, and a meeting half the room missed. Ask in plain language. OSAgent assembles the answer and shows you what it was built from.

Every answer has receipts.

The answer

Not a summary. A claim, with the sources it came from.

OSAgent writes the answer, so the answer sits on its own surface. What it cites sits under it, each source named, so you can go and check any part of it in the tool it came from.

  • The claim is written by OSAgent — the surface says so.
  • The confidence is separate from the authorship, and it is a rail, never a colour on the text.
  • The citations name real sources in your workspace, not a count.
Exampleillustration — synthetic workspace
osagent · ask

Question

Why is billing retry blocked?

checkingSlackGitHubMeetingsCI

OSAgent

High confidence

Billing retry is blocked on the payments-webhook contract. The pull request has been open since Jun 11 with CI failing, and the contract change it depends on has not been merged.

cited from · 3 sources

PR #1284#billing · Jun 12Jun 10 product sync

every cited source opens in the tool it came from

The refusal

If OSAgent can't cite it, it says so instead of guessing.

This is the part of the product we are proudest of, and it is the part that is hardest to fake. Anything can produce a paragraph. The question is what it does when the workspace does not actually hold the answer.

Exampleillustration — synthetic workspace
osagent · ask

Question

What did we decide about enterprise pricing?

checkingSlackDocsMeetings

OSAgent

0 sources

Nothing in workspace memory supports an answer here.

Connect the source that would hold it, or ask something the record covers.

no rail, because no confidence is being claimed

It costs us the demo

A refusal is the least impressive thing a product can do on stage. It is also the only reason to believe the other answers.

It is the same rule every time

Not a confidence threshold you tune until the answers look good. No anchoring evidence, no claim.

It tells you what to fix

A refusal is a map: it names the gap, so you connect the source that closes it.

An answer you can't check is a guess with better grammar.

How it works

Four steps, and the last one can end in nothing.

  1. 01

    The question, as you'd say it

    No query language, no filters, no channel to remember. "Why is billing retry blocked?" is the whole interface.

  2. 02

    Sources resolved

    OSAgent works out which of your connected sources could hold the answer, and retrieves from those.

  3. 03

    Evidence selected

    Retrieved material is only a candidate. What survives is what actually anchors to the subject you asked about.

  4. 04

    Answer, or refusal

    If the surviving evidence supports a claim, you get the claim and the evidence. If it doesn't, you get the refusal.

What changes for you

Stop reconstructing the week to answer one question.

The question “what happened with this?” currently costs somebody twenty minutes of scrolling, two interruptions, and a guess at the end of it.

It becomes a sentence you type, and an answer you can hand to someone else without hedging — because the sources are attached to it.

And when the answer isn't there, you find that out in five seconds instead of at the end of the twenty minutes.

Questions it is built for

  • Why is billing retry blocked?
  • What did we decide about the migration, and who decided it?
  • What changed on this project since last Thursday?
  • Who owns the rollout, and how do we know?
  • Why did the launch date move?

Every one of them is answered from sources your team already produced. Nobody files anything for OSAgent's benefit.

Why you can trust it

The rules the answer is bound by.

Never invent
Ownership, decisions and dates are extracted from real sources. Low confidence is flagged, not smoothed over.
Traceable or absent
A fact that cannot be traced to a source does not enter memory, so it can never turn up in an answer.
It never modifies your work
Asking a question reads your workspace. It never modifies your work in Slack, GitHub or anywhere else.
Scoped to your workspace
Answers are built only from sources your workspace connected, and your data is never used to train models.

What's next

Bring a question you already had to answer the hard way.

OSAgent is in a private pilot. The most useful thing you can do in a walkthrough is ask the question that cost you an afternoon last month — including the ones we should refuse.