Opulent, the Autonomous Knowledge Worker

Give it a whole job in the systems you already run. It comes back finished, evidence attached.

Finance OperationsJuly invoice match

Match every July invoice to its PO and receipt. Hold anything outside tolerance.

Used playbook: Invoice tolerance

I'll match all 412 July invoices against their purchase orders and goods receipts, then hold anything outside your two percent tolerance. Starting now.

Worked for 12m 04s412 matched · 6 held

Three-way match before tolerance. Pull the invoice, its purchase order, and the goods receipt, then compare quantity and price separately so a quantity break does not read as a price break.

heldINV-8841+4.0%$12,480.00$12,000.00

heldINV-8907+4.0%$3,120.00$3,000.00

Done. 412 matched, six held on price, none on quantity. Full report attached.

invoice-exceptions-july.mdOpulent is waiting on your decision

Assign a job with a clear finish

runs on

use cases

Every team runs on triage.

A week has forty hours, so the top twenty accounts get worked and the other hundred and eighty do not. Every tool you bought automates a slice of the work, and none of them automate the person in the middle, moving numbers between screens and chasing what does not match. Your engineers escaped that years ago, when handing work off got safe enough that the list stopped being cut. Opulent brings that to the rest of the company, without asking anyone to work somewhere new. In one engagement, intake that took three to five days finished in under five minutes. Customer-supplied timing, validating in production.

01GTM Engineering
02Finance Operations
03Software Factory
04Research Quality Operations
05Tax Strategy Management
06Autonomous Talent Engineering

Explore all workflows

GTM Engineering

From market signal to verified CRM record.

Two hundred accounts, twenty get worked. Opulent scores all of them, writes the ones that clear your bar, and reads each record back. Freightwave Systems: 78/100, 12/12 fields verified.

Score the inbound list. Write the ones that clear 60 and read each record back.

Task completedScore 200 accounts against the written criteria

Research all 200 accounts

Score against the written bar

Write the ones at 60 or higher

Read every record back

Scored all 200 against your written criteria and wrote only the ones at 60 or higher.

Freightwave Systems, 78/100 · written
Northwind Logistics, 71/100 · written
Cedar Freight Co, 44/100 · below the bar

12 cleared the bar and went to HubSpot. I read every record back: 12 of 12 fields matched.

Finance Operations

From invoice to paid, held, or disputed.

The small exceptions get waved through because nobody has time to chase them. Opulent matches every invoice to its PO and receipt, and holds what breaks tolerance, so fifteen minutes of digging becomes a thirty-second decision. INV-8841: $12,480 against a $12,000 PO. Held, routed, logged.

Match every July invoice to its PO and receipt. Hold anything outside tolerance.

Readvendor-master.csv

Ran command: match

$ match --invoices july --against po,receipt --tolerance 0.02

412 matched · 6 held

Edited invoice-exceptions-july.md

Matched all 412 July invoices in 12m 04s. Quantity matched, 240 of 240. Price did not.

INV-8841, $12,480.00 against a $12,000.00 PO, a 4.0% variance on a 2.0% tolerance. Held, routed with the match attached.

Software Factory

From issue to verified change.

The fix everyone agreed on and nobody scheduled. Opulent makes the change, runs the tests, and watches the rollout. 2.1% errors at 310 ms p99, rolled back before anyone was paged.

Triage the issue, make the change, run the tests, and watch the rollout.

Edited handler.ts

Ran command: pnpm

$ pnpm test --filter webhooks && pnpm soak --requests 1000

passed

8 files changed, +23 -16. Soak passed at 1,000 requests.

Error rate reached 2.1% at 310 ms p99. I marked release 2.4.1 degraded and rolled it back before anyone was paged.

Research Quality Ops

From open question to defensible answer.

The claims that look safe never get checked. Opulent traces all 34 to their originals and holds what the evidence will not carry. 32 of 34 verified, two held back.

Check every claim in the draft. Flag anything a source cannot carry.

Fetchedbls.gov

plan-working.md

Recheck the two unsupported claims

No plan summary provided.

Traced all 34 claims to their originals and verified 32 of 34 quotations.

CLAIM-04, quotation matches the original
CLAIM-19, the source does not say this
CLAIM-27, no primary found

Two claims are held back. Nothing ships without a source.

draft-v3.md18.0 KB

sources.bib8.9 KB

Tax Strategy

From source documents to expert review.

Only the big positions get real scrutiny. Opulent traces every one and marks each assumption where it enters. 41 documents read, two questions routed to your expert.

Trace the position end to end. Mark every assumption and model the alternatives.

Created position-memo-q3.md

+2

Assembled 41 documents and marked 9 assumptions where each one enters.

Treatment A, modeled
Treatment B, modeled
Treatment C, ruled out

Q-04 and Q-07 are open and routed to the reviewing expert. Every figure carries the document it came from. The judgment stays with a person.

transfer-pricing-memo.pdf392.6 KB

schedule-c-2026.xlsx85.9 KB

Talent Engineering

From market map to interview slate.

The candidate on page four never gets read. Opulent maps the whole market and returns dossiers with a source on every claim. 312 sourced, screened to a slate of 9.

Map the whole market for this role and return dossiers ready for a first call.

Agent completedMap the staff engineer market from public record

Agent completedScreen 312 candidates against the written bar

Skilldossier-builder

Command output312 sourced · 48 screened · 9 slated

Sourced 312 from public record, screened 48 against your written bar, and built 9 dossiers.

Every claim carries its source. The candidate on page four got read.

Under every run

Autonomy you can actually grant.

How the work reaches it, what it costs, what it can touch, and how it gets better without anyone rewriting the prompt.

Triggers in useor

Slack message#ops-alerts

Linear issuelabel added

Schedulemonth end, 06:00

Webhookpager, monitoring

Invocation cap40 / hour

Event-Driven Automations

A Slack message, a Linear issue, a GitHub event, a schedule, or a webhook from your pager or your monitoring. The event arrives with the run, so it starts knowing what caused it. An invocation cap keeps a noisy channel from starting forty of them. Nobody prompts it. It starts where the work already happens, and the queue is shorter before anyone opens a laptop.

Reached fromone run

Slacktag it in a thread

Linear, Jiraassign an issue

Emailforward the thread

Webhook, APIanything else

Same Session, same history

Universal Surface

Tag it in a Slack thread, assign it a Linear or Jira issue, forward it an email, or call it over a webhook. Every route reaches the same Session with the same history, so the work carries on when you switch windows. When it needs a decision, it asks in the thread it started in. The surface changes. The worker does not.

Routed by taskper step

Read the ledgersmall

Match and reconcilesmall

Judge the exceptionfrontier

Cost per run, as playbooks growfalling

Frontier Performance

Best cost per execution, because a mature run is mostly not model calls: settled steps harden into checks and code, and frontier reasoning is saved for the decisions that need it. Cost falls as playbooks mature, accuracy holds, and an auditor can follow the trail. A model can tell you how it would do the job. This is what it takes to go and do it.

vendor-portal.example.govlinux

No API. Driven through the screen.

Signed inFiled the returnCaptured the receipt

Recording returned as the proof

computer use

Computer Use

Its own computer: browser, terminal, and a full desktop, so the vendor portal nobody ever integrated is still in reach. It drives the flow the way a person would and hands back the recording, without asking you to migrate anything first. Linux and Windows, with state that survives the waiting.

Run, review, promoteladder

Ran twiceprompt kept

Promoted to Knowledgeretrieved on trigger

Written as a SkillSKILL.md, committed

Shared as a Playbookawaiting your approval

run, review, promote

Self-Improving

Correct it once and it holds for that run. Correct it twice and the context is worth keeping. A standing fact becomes Knowledge, a repeated procedure becomes a Skill, and recurring work becomes an Automation. A one-time exception stays a one-time exception. Nothing is promoted until a person approves it, which is the only reason the learning is safe to trust.

Connected systems

It works inside your stack, not beside it. (1)

Access is granted per system and per action, and reading is granted separately from writing. Before anything consequential, the run stops and asks the person who owns it. What it writes lands in your system of record. The systems with no API are still in reach. Opulent drives those through the screen, the way a person would.

start here

Start with one job.

Pick one from the library, connect the systems it touches, and write down what you expect to inspect at the end. Then run it against real work this week. The first run won't be perfect. That's what the loop is for.

Deal sourcing Invoice exceptions Release verification

Try Opulent Request a demo →

Or read a completed run first

Opulent, the Autonomous Knowledge Worker

Every team runs on triage.

From market signal to verified CRM record.

From invoice to paid, held, or disputed.

From issue to verified change.

From open question to defensible answer.

From source documents to expert review.

From market map to interview slate.

Capacity you can't hire for.

Ship the backlog you shelved

Repair the whole pipeline, not the top of it

See the renewal risk before the renewal

Close with every exception already chased

Every launch fully shipped

Read every contract, not a sample

Catch the signal before it becomes a fire

Hire from the whole market

Every customer signal lands in the spec