Academy

What it takes to trust it

Trust is the top rung because it is the hardest — authority earned gradually, every action traceable to its reason, and a system that survives an audit while running in an isolated managed environment.

The Felesh team 4 min

You have watched an agentic system handle real work for six months. It has acted on your behalf, coordinated specialists, and worked through your documents. It has been right more often than it has been wrong. Now someone in your organisation asks: could we give it more authority?

A track record built from receiptsThree dated actions from six months of work, each ticked and each leaving a receipt: in February it sent refund #4127 on your behalf, in April it passed ticket #9310 to the billing team, in June it matched invoice INV-889 to order PO-55. Each receipt slides into a track record where every action sits with its reason. From that record comes the question: could we give it more authority?six months of work, with receiptsFeb · on your behalfrefund #4127 sentApr · with specialiststicket #9310 to billingJun · in your documentsINV-889 matched to PO-55track recordaction + reasoncould we give it more authority?
Six months of actions, each one leaving a receipt, build the track record that makes someone ask for more authority.

The question carries weight. Not because the system is unreliable, but because authority at scale is a different category of risk from authority in a single task. The moment you consider expanding what the system can initiate, you want to know: can I see the receipts for every decision? Can I understand the reasoning? Could a regulator, an auditor, or an unhappy client follow the trail?

Following the trail of one decisionThe action, refund #4127 for 1.2M, is linked to its receipt: made by the billing agent on 14 June at 10:42, under the rule for refunds under 2M, citing invoice INV-889. The receipt is linked to the reasoning: INV-889 was charged twice. A trail runs from the action through the receipt to the reasoning, and a regulator, an auditor or a client can follow it.the actionrefund #4127amount 1.2Mreceiptby billing agentat 14 Jun 10:42rule: under 2Mcites INV-889reasoningINV-889 wascharged twicea regulator, an auditor or a client can follow it
One refund leaves a receipt that points to its reasoning, so a regulator, an auditor or a client can follow the trail end to end.

Trust is not a feature you configure. It is earned action by action, receipt by receipt, over time that cannot be compressed. The four properties we build towards at this stage: authority that grows along a defined trajectory via the autonomy dial; every action attached to its reasoning so no decision is opaque; a glass-box review record that survives external scrutiny; and strict per-organisation isolation.

The autonomy dial, notch by notchA table for one action, refund #4127 for 1.2M, on the autonomy dial. Drafts only, from day one: it drafts the refund and a person sends it. Suggests, you approve: earned, for example, after 300 receipts. Acts, reports after: earned, for example, after 1,100 receipts. Acts alone, on refunds under 2M, is not yet earned. Each notch is earned at the one below and can be turned back.dialwhat it may do with refund #4127 · 1.2Mearned bydrafts only: a person sendsday onesuggests, you approvee.g. 300 receiptsacts, reports aftere.g. 1,100 receiptsacts alone, refunds under 2Mnot yeteach notch is earned at the one below, and can be turned back
On the autonomy dial each notch is earned by receipts at the notch below, and it can be turned back.

One direction we are working towards: a system that learns from its own operational history — what we call a self-learning colleague. We want to be clear: this is a roadmap direction, not something we have shipped. We share it because it shapes the architecture decisions we are making now.

Shipped history, roadmap colleagueOn the shipped side, an operational history fills with dated lines: refund #4127 approved, ticket #9310 escalated, INV-889 flagged, and people read it. On the roadmap side, behind a dashed line, a self-learning colleague would learn from that history. It is marked not shipped.shipped todayroadmapoperational history14 Jun refund #4127 approved15 Jun ticket #9310 escalated16 Jun INV-889 flagged twicerecorded now, read by peoplewould learn from itself-learningcolleaguenot shipped
The operational history is recorded today; a colleague that learns from it is a roadmap direction, not something shipped.

The four reads below ground each of these properties: