Academy

Authority is a trajectory, not a key

You don't hand an intern the keys to everything on day one. AI agents should earn their authority along a path too — just like a career.

The Felesh team 4 min

Picture a new intern walking into an organisation on their first day. You hand them small tasks, supervised. Once they handle those well, larger ones. Months later, when they have shown again and again that they can be relied upon, you hand them real authority. No one gives an intern the keys to the vault on day one. Authority is not granted all at once; it is earned along a path.

We believe AI agents should travel the same path. An agent’s authority — what it is allowed to do without oversight — should not be a fixed setting at the moment of installation. It should rise over time, backed by evidence. Authority is not a toggle flipped at the start; it is a trajectory the agent walks.

A fixed setting versus a pathTwo panels. Left, a toggle set once at install: store credit of 300,000 is allowed from day one to month six, with no record behind it. Right, a path walked over time: the agent first tags tickets, then drafts replies, and only by month six issues credit of 300,000, each rise backed by a record. Same top, reached with evidence.a toggle, set at installa path, walked over timeon, for everythingcredit 300,000no record behind itday 1month 6tagsdraftscredit 300,000day 1month 6same top, with a record behind each rise
A fixed setting hands over everything on day one; a path reaches the same authority only with a record behind each rise.

From the bottom of the path, upward

Every path has an entry point. For a new agent, this is the lowest-risk work: tasks where the cost of correction is negligible and a human operator can easily review the output. This is the bottom of the path — where the agent shows its work and its behaviour is measured.

As it climbs, two things change at once: the risk of the task grows, and human oversight shrinks. This progression is neither automatic nor time-based; it is conditional. An agent moves up a step only when it has repeatedly and stably completed the step below. The path runs upward, but every step is earned.

The climb, step by stepA staircase of four steps for the same support agent. Step one tags ticket 7731 as shipping, where a mistake is cheap to fix. Step two drafts the reply. Step three sends the reply and a person reads it after. Step four issues store credit of 300,000. Risk grows from the first step to the last while human review shrinks, and each step is reached only after a stable run on the one below.each step up: a stable run on the step belowcheap to fixtags #7731as shippingdrafts thereply #7731sends replyread aftercredit300,000riskreview
Each step up carries more risk and less review, and the agent climbs only after a stable run on the step below.

Evidence, not promises

This differs entirely from “trust the AI.” Trust is not assumed; it is built. Every step is backed by evaluation: real examples demonstrating whether, at a specific level of responsibility, the agent’s behaviour is stable and inspectable. Increased authority requires increased evidence.

Evidence unlocks the next stepAn evidence sheet for step two, drafting replies: five real tickets, 7728 to 7732, each draft reviewed and ticked one after another. When the sheet is stable, the lock on step three, sending the reply, opens. Each draft stays open to inspection.evidence at step 2: drafts the reply7728 parcel late7729 wrong size7730 address typo7731 order status7732 cancel orderreviewedreviewedreviewedreviewedreviewedstep 3sends the replylockedunlockedopens when stablereal tickets, each draft open to inspect
The next step opens on a sheet of real examples, each one open to inspection.

Because the path is evidence-based, it is also reversible. If an agent’s behaviour degrades — perhaps the underlying data has shifted or the workflow has grown more complex — its authority is lowered. This is not a failure; it is the mark of a healthy system. Just as responsibilities in a functional organisation are tuned to performance, an agent’s authority remains dynamic and accountable.

Authority up and down with evidenceA line chart of one agent’s authority from March to July across three notches: suggests, you approve; acts, reports after; acts alone. It starts at the lowest notch, rises to the middle after a stable run, rises to acts alone in June, and in July, when the data shifts and behaviour degrades, steps back down to the middle notch. Authority follows the latest evidence, up or down.acts aloneacts, reports aftersuggests, you approveMarAprMayJunJulstable runstabledata shiftedauthority follows the latest evidence, up or down
Authority climbs with evidence and steps back down when behaviour degrades.

A forward-looking idea

Part of what we describe here is the direction of our design, not the claim of a finished product. The goal is clear: systems that manage authority like a path, not like an on/off switch. This view moves AI from a static tool toward something closer to a colleague that grows over time and earns trust.

Don’t gift authority on day one. Let the agent earn it along the path, the way any professional does. You read a page from the bottom upward; you climb a career step by step; and the authority of a trustworthy agent walks that same upward path.