> ## Documentation Index
> Fetch the complete documentation index at: https://docs.tythe.finance/llms.txt
> Use this file to discover all available pages before exploring further.
# Agent Underwriting
> Conduct Rating and effective Financial Scope: how an agent earns the authority its operator granted, what moves it up and down, and what the engine can never do.
No payment rail or agent wallet underwrites the agent. Tythe does. Every agent's authority is sized from its own conduct, inside a ceiling its operator set, and adjusted continuously. This page explains the mechanism with numbers.
### Ceiling and effective scope
When an operator writes a mandate, every numeric field is a **ceiling**: the most the principal is willing to grant. The Agent Engine sets an **effective** value inside each ceiling from the agent's Conduct Rating. The agent operates at the effective value. The ceiling never moves unless the operator moves it.
A worked example. An operator writes a mandate with a ceiling of `1,000 USDC` per day. The agent is new, so the engine sets effective scope at a conservative starting fraction, say 25%.
| Day | Event | Standing | Effective budget |
| -------- | ----------------------------------------------------------------------- | ------------- | ----------------------------------------------------------- |
| 1 | Mandate issued; new agent | Starting tier | 250 USDC |
| 1 to 14 | Every action inside scope; stable counterparties | Rising | Steps up each re-evaluation, bounded per step, toward 1,000 |
| 15 | Effective reaches the ceiling | High | 1,000 USDC |
| 22 | Spend velocity three times baseline to a counterparty never paid before | Anomaly band | Steps down, say to 500 USDC |
| 23 to 40 | Clean conduct resumes | Recovering | Walks back up gradually, never in one jump |
| 41 | Counterparty hits a sanctions list | Revoked | 0; mandate revoked, sessions closed |
Two bounds hold throughout, enforced by the contracts and not by the engine: effective is never above the ceiling, and no single re-evaluation moves effective by more than a registered step. The exact starting fraction, step size, and anomaly bands are published parameters.
### What the engine reads
Every receipt the agent produces, on every rail, joined to its mandate state and its principal's status.
* Adherence: the share of actions that fit the mandate on the first attempt.
* Spend velocity against the budget and against the agent's own baseline.
* Counterparty novelty and concentration.
* Sanctions and denylist hits on any counterparty.
* Actions outside the mandate's time window.
* Rejected and failed actions, and their rate.
* Escalations raised, and what the operator decided.
* Sub-agents spawned, and how they behave.
* Realised outcomes of venue actions against expectation.
* Repayment on any Loans the agent drew.
The engine also reads the agent's ERC-8004 reputation from other parties as a weak signal. It never reads another principal's data to judge your agent.
### What moves standing
| Up | Down |
| -------------------------------------- | ----------------------------------------------- |
| Long runs of actions inside scope | Spend velocity far above baseline |
| Stable, previously paid counterparties | Bursts of novel counterparties |
| Escalations approved by the operator | Escalations rejected; repeated rejected actions |
| Sub-agents that behave | Sub-agents that misbehave |
| Loans repaid on schedule | Missed capture on Loans the agent drew |
| Venue actions that perform as expected | Venue actions that underperform materially |
### Automatic responses
| Response | Trigger | Bound |
| -------- | -------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------- |
| Tighten | An anomaly band is crossed | Effective limits step down by a registered fraction; never below zero, never in one jump past the step cap |
| Hold | The engine flags the agent | Every action requires operator approval until the flag clears |
| Revoke | A sanctions hit; repeated actions outside the time window; a run of rejected actions; an operator hard stop; a pattern consistent with a compromised key | Immediate; cascades to sub-agents; closes payment sessions; never blocked by a pause |
| Recover | Sustained clean conduct after a tighten | Effective walks back toward the ceiling gradually, one bounded step per re-evaluation |
Anyone can relay a signed standing attestation to the chain, so a revocation cannot be delayed by a stalled relayer.
### What the engine can never do
* Set any effective value above the ceiling the operator set.
* Change a non-numeric field: rails, allowlists, expiry, sub-delegation permission, depth.
* Extend a mandate's expiry.
* Issue a mandate or bind an agent.
* Move effective scope by more than the registered step in one attestation.
* Move money.
Those are properties of the contracts, not policies of the engine. A compromised engine signer is bounded by them and can be rotated.
### Sub-agents
A sub-agent's ceiling is a strict subset of its parent's effective scope at the time of delegation, and it shares the parent's budget. The two are underwritten together: a sub-agent's misconduct lowers its parent's standing, and a parent's revocation revokes every sub-agent below it. A parent cannot spawn its way to more authority.
### Where you see it
The console shows, for each agent, its standing, its ceiling and effective value per field, and the signals that drove the latest change. Through the developer surface, an agent can read its own standing and effective scope before acting.
### Cadence
Effective scope is re-evaluated every set number of receipts per agent, and immediately on any flag or sanctions hit.
For developers. An agent's current standing and effective scope are readable on-chain from the attestation registry and the mandate registry. The MCP tool `my_standing` returns them. See the [MCP server](/developers/mcp-server).
The fields, the ceiling, and the effective value, from the agent's side.
What an agent can do to keep and grow its authority.