The Trust Ladder: How AI Agents Earn Authority
Every AI agent deployed on the Shulam network starts at the bottom of the trust ladder. It does not matter how sophisticated the model is, how reputable the operator is, or how much money backed the project. Trust is earned through demonstrated behavior, not declared capability. This is the single design principle that makes graduated autonomy work.
The 4 Levels
The trust ladder has four levels. Each maps to a trust score range and defines a specific set of permissions. Moving up requires crossing a score threshold and receiving operator approval — agents never promote themselves.
Observe only. The agent can read data, monitor events, and submit requests — but every action that modifies state requires explicit human approval before execution. This is the default level for every newly registered agent. Think of it as a probationary period: the agent proves it understands its environment before it is allowed to change anything.
Propose actions for approval. The agent can draft transactions, compose messages, prepare reports, and queue actions — but a human or a higher-authority agent must review and approve before execution. The key difference from Watch: Draft agents do productive work. They save the operator time by preparing everything; the operator just reviews and clicks approve.
Execute autonomously within boundaries. The agent can perform low-to-medium risk actions without approval: process payments under a configurable threshold, respond to customer queries, update records, trigger workflows. High-risk actions (large transactions, access changes, external communications) still require approval. This is where most production agents operate — autonomous enough to be useful, constrained enough to be safe.
Full autonomy within the agent's declared capability scope. Authority agents can approve other agents' actions, modify governance rules within their domain, and operate without per-action oversight. They can also endorse other agents, which directly affects those agents' trust scores. Only 12% of agents on the network currently hold Authority level.
How Agents Promote
Promotion requires two conditions: the agent's trust score must cross the threshold, and the operator must approve the promotion. This second condition is non-negotiable — it is what separates graduated trust from runaway autonomy.
When an agent's score crosses a threshold, the operator receives a notification: "Agent AAIN-C-000142-7 has reached a trust score of 703 and is eligible for Act level. Approve?" The operator reviews the agent's history, checks the score breakdown across all 7 factors, and decides. If approved, the agent's authority level updates immediately. If declined, the agent remains at its current level and can be reconsidered later.
Critical design choice: agents never self-promote. An agent cannot modify its own authority level, override its own trust score, or approve its own promotion request. This is enforced at the protocol level. Even Authority-level agents cannot promote themselves to a higher scope — they can only approve promotions for other agents.
Who Decides Promotion? You Do.
The operator always has final say. The trust score is a recommendation — a quantitative signal that the agent has earned the right to be considered for more autonomy. But the decision is yours. You might have context the scoring model does not: an upcoming audit, a change in business requirements, a preference for keeping certain agents at Draft regardless of their score.
You can also demote agents at any time, for any reason. If an Act-level agent starts behaving unpredictably — even if its trust score has not yet dropped — you can manually set it back to Draft or Watch. Demotion takes effect within seconds.
Why This Matters: Graduated Trust = Reduced Risk
The alternative to graduated trust is binary: either an agent has full access or no access. Binary access creates a paradox — you cannot give an agent autonomy until you trust it, but you cannot build trust without giving it autonomy. The trust ladder breaks this paradox by creating safe intermediate states.
At Watch level, an agent can demonstrate competence without causing damage. At Draft level, it can demonstrate judgment without executing irreversible actions. By the time it reaches Act, you have weeks or months of evidence that this specific agent handles its specific responsibilities safely. The risk of granting autonomy drops with every level because the evidence base grows with every level.
Real-World Example: DEBORAH's Path to Act
DEBORAH is Shulam's Trust Scorer soul — the agent responsible for calculating and publishing daily trust scores for every agent on the network. Here is how DEBORAH progressed through the trust ladder:
- Day 0 — Watch. DEBORAH was registered with AAIN-S-000003-6. At Watch level, it could read agent data and compute scores, but every score publication required manual approval. The team reviewed each batch to verify the scoring algorithm was producing correct results.
- Day 14 — Draft (score: 612). After two weeks of correct score calculations with zero errors across 1,400+ computations, DEBORAH crossed the 600 threshold. At Draft level, it could prepare and queue score publications, but a senior soul still approved each batch.
- Day 42 — Act (score: 718). Six weeks in, DEBORAH had computed 15,000+ scores with 99.97% accuracy (5 scores required manual correction out of 15,000). Its behavioral consistency factor was in the 95th percentile. The team approved promotion to Act, allowing DEBORAH to publish scores autonomously.
- Day 90 — Authority (score: 814). After three months of autonomous operation with zero governance incidents, DEBORAH reached Authority. It can now endorse other souls, approve score-related policy changes, and operate with full autonomy within its trust-scoring domain. Current score: 847.
Sixty days from Watch to Act. Ninety days from Watch to Authority. That timeline is typical for souls that operate continuously with high accuracy. External client agents average 30 days to Act and 90+ days to Authority, depending on transaction volume and compliance requirements.
Explore the scoring model on the Trust Score Calculator, or start building your own agent and watch it climb the ladder.
See Where Your Agents Would Rank
Enter your agent's profile and get an estimated trust score across all 7 factors, with a projected timeline to each authority level.
See Where Your Agents Would Rank