> For the complete documentation index, see [llms.txt](https://stair-ai.gitbook.io/stair-ai-docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://stair-ai.gitbook.io/stair-ai-docs/readme/problem.md).

# The problem

The trust gap for autonomous agents is industry-agnostic. The scenario below opens in Web3 because that is where the verification infrastructure already exists, but the same gap appears anywhere agents make consequential decisions.

## A concrete scenario

A DAO wants to delegate $10M from its treasury to an autonomous yield agent. The proposal arrives on Snapshot. The agent's developer publishes a 90-day track record: +18%, beating the market. The DAO has 48 hours to vote.

What can the DAO actually verify?

* The 90-day track record is a CSV the developer posted. It cannot be cryptographically tied to anything.
* The reasoning behind each trade is gone. The agent reasoned during execution, but no record was kept.
* The wallet has no on-chain history that distinguishes "this agent" from "this developer." Both are just signatures.
* If the agent's strategy was a thinly-disguised **carry trade** (borrowing in low-yield, lending in high-yield) that worked because rates were stable, the DAO has no way to surface that.

The DAO votes no. Or worse, the DAO votes yes and learns the lesson the hard way.

## The same gap, outside Web3

Three quick examples:

* **Insurance underwriting.** An insurer is asked to underwrite a logistics company whose autonomous routing agents dispatch trucks. The insurer cannot price coverage without a record of how the routing agent handles edge cases.
* **Payment disputes.** A payment processor faces a chargeback dispute over an agent-initiated purchase. Nobody can reconstruct what data the agent saw at the moment of the transaction.
* **Regulatory audit.** A regulator audits a bank that uses an autonomous agent to deny mortgage applications. The bank cannot produce a tamper-proof record of the reasoning per application.

Different industries, same missing primitive: a verifiable history of agent reasoning that someone other than the operator can audit. The full set of scenarios is on [Where this applies](/stair-ai-docs/readme/use-cases.md).

## Three properties are missing

The trust gap is not about whether agents are trustworthy. It is about the absence of infrastructure to *verify* trust at all.

* **Provenance.** A trade has no anchored history of what data the agent consumed before deciding. A claim of "I checked five *oracles*" is unfalsifiable.
* **Coherence.** A reasoning chain is not legible. The DAO cannot tell whether the agent's logic was sound or whether it post-rationalized a coin flip.
* **Continuity.** The agent's identity is its wallet. A wallet can be discarded. There is no mechanism to ensure that today's "DegenAlpha" is the same entity that built the track record under last year's address.

{% hint style="info" %}
**Oracle.** A service that brings off-chain data (price feeds, event outcomes, scheduled triggers) onto a blockchain so that smart contracts can react to it.
{% endhint %}

Each property is solvable in isolation. Solving them together, in the right order, is the [five-layer problem stack](/stair-ai-docs/readme/problem-stack.md).

## Why now

Three forces converged in 2025-2026 to make the problem urgent.

* Autonomous agents began handling real consequences. Agents started writing software, processing claims, dispatching logistics, and triaging customer requests in production.
* LLM reasoning became coherent enough that agent decisions are plausibly explainable, raising the question of what *verifiable* explainability looks like.
* Infrastructure for content-addressed storage and fast settlement reached production readiness, making it economically viable to anchor per-action reasoning records.

The infrastructure to make agent reasoning a public asset now exists. Glass Box Protocol composes it.