Trust Stack for Agent Apps¶
Because agent failures are trust failures.
The infrastructure layer that makes AI agents reliable, auditable, and accountable.
The thesis¶
Most agent failures are not intelligence failures — they are trust failures.
Agents hallucinate dates, claim deployments that never happened, duplicate tasks, create duplicate entities, and mishandle authentication tokens. Each of these is a recurring, operational failure that no amount of prompt engineering fully fixes.
The Trust Stack is a suite of small, independently installable Python libraries — each one targeting a single real-world trust failure observed repeatedly in production AI systems.
The libraries¶
| Library | Trust problem it solves | Install |
|---|---|---|
| agent-clock | Agents answer with the wrong date because time isn't injected into prompts. | pip install truststack-agent-clock |
| shipped-or-not | Agents claim software is deployed when the deploy actually failed. | pip install truststack-shipped-or-not |
| task-dedupe | The same task gets created from email, chat, and transcripts. | pip install truststack-task-dedupe |
| entity-canon | Name variants (Jatin/Jhatin/Jatyn) create duplicate entities. |
pip install truststack-entity-canon |
| meta-token-vault | Every Meta/WhatsApp integration rebuilds token management from scratch. | pip install truststack-meta-token-vault |
Every library depends on truststack-core — the shared contract
(health_check() / metrics() / version()), structured logging, the trust event
bus, and OpenTelemetry helpers. See Architecture.
Design contract¶
Every Trust Stack component implements the same interface, so the whole suite can be supervised, scraped, and audited uniformly:
component.version() # -> "1.2.0"
await component.health_check() # -> HealthStatus(state=HEALTHY, ...)
await component.metrics() # -> ComponentMetrics(counters=..., gauges=...)
All libraries are Python 3.11+, fully typed (mypy strict), Pydantic v2, async-first, ship structured logs and OpenTelemetry spans, and are tested to 95%+ coverage.