Skip to content
Get started

Help your agents adapt and act proactively.

Join the waitlist
  • SupportTickets · 94% closed, no review
  • FinanceMonth-end close · done in 2 days
  • EngineeringOn-call · rollbacks in 4 minutes
  • OperationsRobot fleet · picks holding at 96%

Scale autonomy safely across every agent deployment.

Help your agents adapt across the business, act proactively and take on more work within the limits your team sets.

Join the waitlistExplore the developer tools

Finished in the last hour

  • SupportRefund issued inside policy
  • On-callRolled back 3f9c2, p95 at 180ms
  • Finance212 invoices matched, 3 held
  • ITAccess ready for 9 new hires
  • RevenueNorthwind renewed at 10%
  • RoboticsRobots 3 and 7 rerouted
  • ComplianceEvidence pack sent to audit
  • Billing38 failed payments recovered
  • EngineeringFlaky test fixed, PR opened
  • CollectionsHalden and Corvo paid

What your agents get.

From the weekly report Self keeps on every agent: measured on your own work, in your own systems, against a standard you can check.

Better than the benchmark, on your own work.

What’s measured
The share of tasks done right, checked in the system each task was done in.
Compared with
The same model’s score on the public benchmark for that kind of task, and the agent’s own first week.

Tasks done right, by workflowWeek 12 of 12

+22points

against the model’s public benchmark, averaged over five workflows, in week 12.

  1. Support ticketsZendeskSupport tickets: 64% in week 1, 94% in week 12; the public benchmark is 71%.
  2. Invoice matchingNetSuiteInvoice matching: 77% in week 1, 98% in week 12; the public benchmark is 82%.
  3. Incident triagePagerDutyIncident triage: 55% in week 1, 91% in week 12; the public benchmark is 63%.
  4. Access requestsOktaAccess requests: 72% in week 1, 97% in week 12; the public benchmark is 78%.
  5. Contract reviewIroncladContract review: 58% in week 1, 89% in week 12; the public benchmark is 66%.

Every finished task costs less than the week before.

What’s measured
All a task costs (the model, its tools, and the minutes people spend reviewing it) divided by the tasks finished and verified.
Compared with
Week 1, the same agents on the same work. It falls because fewer tasks need a review, and the skills it learns run on smaller models.

Cost per finished taskWeek 12 of 12

$0.19

per finished task in week 12, 90% less than in week 1. 31 learned skills now run on smaller models.

  1. Week 1: $1.84 per finished task.
  2. Week 2: $1.68 per finished task.
  3. Week 3: $1.49 per finished task.
  4. Week 4: $1.30 per finished task.
  5. Week 5: $1.07 per finished task.
  6. Week 6: $0.85 per finished task.
  7. Week 7: $0.64 per finished task.
  8. Week 8: $0.47 per finished task.
  9. Week 9: $0.34 per finished task.
  10. Week 10: $0.26 per finished task.
  11. Week 11: $0.21 per finished task.
  12. Week 12: $0.19 per finished task.

More of the work finished with no review, and mistakes kept under your limit.

What’s measured
Tasks finished and verified in your systems with no person checking first, and tasks later reversed or corrected, within 30 days.
Compared with
Week 1, and the error limit you set. It takes on more only where its record stays under that line.

Finished with no reviewWeek 12 of 12

92%

of tasks finished with no person checking first, up from 38% in week 1.

Later found wrong

Finished with no review rose from 38% in week 1 to 92% in week 12. Later found wrong fell from 0.9% to 0.3%, under the 1% limit.

It acts on your numbers before anyone asks.

One Tuesday morning with a support agent on Self.

Refund rateShare of orders refunded in the last hourWatched by the support agent

2.1%Inside its usual rangeTuesday, 1:10 PM

9:40 AMNoticed. No one had asked.9:52 AMDana approved the claim.1:10 PMBack to usual.
  1. 9:40 AMNoticed. No one had asked.
  2. 9:52 AMDana approved the claim.
  3. 1:10 PMBack to usual.
Support leadsDana, Marco, Priya, Support agent
  1. Tuesday
  2. Support agentRefund rate since 8 AMusual2.6%8:009:40Refund rate is 2.6% and climbing. Usual is 2.1%. Looking into it now.9:40 AM
  3. Found it. 38 of the last 52 refunds are one product: Kettle K2, batch 0419. The lids crack in transit.9:43 AM
  4. Inside my limits, I paused auto-refunds on that batch, sent replacement lids instead, and held the 212 still in the warehouse.9:45 AM
  5. Refund brief$12,480to claim from Halden SupplyKettle K238Others14Kettle K2, batch 0419: why refunds jumpeduseself.aiThe brief, if you want the detail.9:46 AM
  6. Dana, one call is yours. Halden Supply owes us for the batch, and a claim over $5,000 needs you.

    Claim $12,480 from Halden Supply?

    Select one

    File the claim1
    Wait for their QA report0

    9:46 AM

    View votes

    9:46 AM
  7. file it. ask them for the QA report too9:52 AM
  8. Support agentFiled, and asked for the report. Refunds are already slowing.9:53 AM
  9. Back to 2.1%. Batch 0419 stays paused until Halden answers.1:10 PM

You draw the lines. It asks before it crosses one.

Every action is checked against your limits before it runs. The record keeps what it did, the limit it met, and who allowed it.

Support agentLive

Limits

Under a line, it acts on its own. Over it, the person you name decides.

  • Refunds up to $300Acts on its own
  • Refunds over $300Dana K. decides
  • Pause a product’s auto-refundsActs, then tells the team
  • Supplier claims up to $5,000Acts on its own
  • Supplier claims over $5,000Dana K. decides

Handed back when anything changes. A new model, policy or tool sends its decisions back to a person until 50 in a row are approved clean.

Record

Every action, the limit it was checked against, and who allowed it.

  1. Thu 4:31 PMRefunded $410 on ticket 49022Refunds over $300Dana K., in WhatsApp
  2. Thu 4:20 PMRefunded $96 on ticket 49017Refunds up to $300Inside the limit
  3. Thu 4:12 PM50 approved clean. Refunds up to $300 are its own again.Autonomy earned backSelf
  4. 10:34 AMRefunded $129 on ticket 48262Refunds up to $300Dana K., in WhatsApp
  5. 10:21 AMRefunded $64 on ticket 48240Refunds up to $300Dana K., in WhatsApp
  6. 10:15 AMModel updated. Refunds go back to Dana until 50 are approved clean.Autonomy handed backSelf
Self for enterprise: scale autonomy safely across every agent deployment