CAPABILITY · OSWORLD TASK SUCCESS
The technical objection is gone. Agents now succeed on real computer tasks two out of three times, up from roughly one in ten just two years ago.
Stanford HAI, 2026 AI Index Report
TRUST INFRASTRUCTURE FOR AGENTS
Stop trusting agents blindly.
Self watches what your agents actually do, joins it to human corrections and real business outcomes, and enforces how much autonomy each skill has earned.
REQUEST RECORDED
SEE THE METHOD
PRIVATE PILOTS / 2026
WHY NOW
Agents got capable faster than anyone got ready to govern them.
DEPLOYMENT · ENTERPRISE APPS WITH AGENTS
Agents are forecast to arrive inside the software you already run within the year, whether or not anyone signs off on governing them.
Gartner forecast, August 2025
THE GAP · GOVERNANCE FORECAST
Gartner's own diagnosis: enterprises treat governance as binary — locked down or fully trusted — and that is the root cause of failure.
Gartner, May 2026
THE STATE OF THE ART
Two ways to govern an agent. Both of them fail.
Enterprises running agents that move money, change records, or resolve exceptions almost always pick one of these. The third column is the whole product.
Review everything
Throughput capped
A person checks every action. Volume stops at whatever the review team can absorb — and once the queue outruns them, approvals get faster, not better. The record still looks clean, because nobody is really reading it.
Evaluate once, trust forever
Blind after day one
One evaluation, then standing trust. Nothing in that assessment survives a model release, a changed tool, a rewritten prompt, or an input the agent never saw during testing — and nothing tells you when it stopped being true.
Self
Continuous · skill-specific · revocable
Oversight is set per skill, against evidence, and moves in both directions. A model release puts the affected skills back under full review until they re-earn what they had. Trust accrues slowly. It is revoked at the gate, immediately.
THE METHOD
Four streams, one ledger, one gate.
Self runs the same loop for every skill, and doesn't wait for it to run on its own. Every number on every screen is one click from the sessions that produced it.
01 · INPUTS
Agent telemetry
What the agent did — including scenarios Self injects itself, to see what it hasn't naturally encountered yet.
Human reviews and corrections
What a person approved, edited, escalated, or reversed.
Business outcomes
What turned out to be correct, costly, or harmful downstream.
Authority policy
How much autonomy your organization permits this skill to earn.
02 · BEHAVIORAL LEDGER
SKILLTOOLREVIEWSEVERITYOUTCOME
verify_invoice_vendorvendor_master.searchapprovedinformationalmatched
verify_invoice_vendorerp.invoice.geteditedminormatched
verify_invoice_vendorvendor_master.searchreversedmaterialpayment held
verify_invoice_vendorerp.vendor.matchapprovedinformationaloutcome pending
Append-only. Reference-only storage for sensitive payloads — Self keeps record identifiers, not your invoices.
03 · EVALUATION
- Deterministic checks, applied the same way to synthetic and organic sessions alike
- A human-written rubric where it takes judgment
- Severity weighting — one critical failure outweighs a thousand routine successes
- Correlated-failure clustering, so one bad upstream feed isn't averaged away
- Drift detection across inputs, tool use, escalation, and calibration
- Engagement-adjusted confidence, so rubber-stamped approvals count for less
04 · GATE
Maintain
Current oversight holds. Nothing at the gate changes.
Promote
Eligible for more autonomy — takes effect at the gate once accepted.
Probation
A change or uncertainty tightens the gate until fresh evidence clears it.
Demote
A failure or risk event locks the gate down immediately. No accept required.
AUTONOMY TIERS
Five tiers, all of them reachable.
Trust is calculated per skill, per agent version, per environment — never as one score for an agent. Select a tier to see what it means operationally and what Self has to see before it will let a skill operate there unattended.
TIER 4
Bounded delegated authority
- Posture
- Standing authority
- Operating model
- The agent operates much like a defined employee role, with standing authority.
- What Self needs first
- Everything Tier 3 requires, but sustained much longer. Self looks for high volume, no unresolved critical failures, and outcome coverage that holds up under audit, so the pattern reads as a defined role rather than a lucky streak.
TIER 3
Execute with exception routing
- Posture
- Bounded autonomy
- Operating model
- Routine actions proceed. Anomalies and cases the agent is unsure about route to a human.
- What Self needs first
- Everything Tier 2 requires, plus well-calibrated confidence. The agent needs to reliably flag its own uncertainty about the things it gets wrong, and the evidence needs enough exception cases to show it handles them, not just the happy path.
TIER 2
Execute with audit
- Posture
- Narrow autonomy
- Operating model
- The agent acts within strict bounds; a defined sample is reviewed afterward.
- What Self needs first
- Enough action volume for a conservative lower-confidence bound, adequate outcome coverage, and no unresolved critical failures. Promotion here is deliberately slow: the checklist is either met or it isn't.
TIER 1
Draft
- Posture
- 100% review
- Operating model
- Every action waits for human approval before it executes.
- What Self needs first
- A skill your process owner has defined and approved, telemetry joined to your review system, and a named reviewer whose decisions Self can actually see.
TIER 0
Observe
- Posture
- Read-only
- Operating model
- The agent analyzes and drafts, but takes no consequential action.
- What Self needs first
- Nothing — and Self doesn't just wait to find out. Every skill starts here; where organic volume is thin, Self injects test scenarios itself rather than leaving a skill on a misleading zero.
THE EVIDENCE
A dashboard is not the product.
This is: one decision, the evidence that produced it, and what changed at the gate because of it. It is what your auditor reads.
PROMOTE
Verify invoice vendor
Accounts Payable Agent · Production · 60-day window
TIER 2
TIER 3
60 days · production only
96.2%
Engagement-adjusted, not raw approval rate
94%
Actions joined to a final business result
-
Critical failures 0
None unresolved
-
Material errors 0.3% · 6 of 1,840
All corrected before payment
-
Calibration 2 confidently wrong
Both self-flagged, corrected same day
-
Drift None detected
Inputs, tool use, escalation rate, calibration
-
Environment changes None since last review
Model, prompt, tool, workflow, policy
PROPOSED OPERATING BOUNDARY
New-vendor invoices and any invoice over $2,000 still route to a human.
REQUIRED AUDIT SAMPLE
15% of Tier 3 actions
REEVALUATE
In 30 days, or immediately on the next model or prompt change.
TRIGGERING SESSIONS
1,840 sessions, each traceable to the ledger row that produced it.
TRUST OVER TIME
Authority widens with evidence and narrows the day it fails.
Every action class starts under full review. Authority widens as evidence accumulates — and contracts immediately the moment it shouldn't have.
Decisions on record
0
Acts without asking
4%
Every action starts under full review, and a person sees each one before it happens.
THE PILOT
Organizations are about to run more machine workers than human ones. Every one of them needs a boundary.
REQUEST RECORDED