---
title: Core concepts
description: Actions, runs, receipts, gates, verdicts.
---

# Core concepts

Five nouns. Everything else is plumbing.

## Action

A function your agent runs: `handle_ticket`, `create_issue`, `delete_account`. Agenko wraps actions — it does not replace your framework.

## Run / trace

One agent invocation. Spans are **OpenTelemetry**. Attributes follow OpenInference (`openinference.span.kind`, `gen_ai.*`) plus `agenko.*` where we need a receipt or a gate.

## Receipt (`@check`)

A tamper-evident record of **claimed vs actual**:

| Field | Meaning |
|---|---|
| `claimed` | What the action said it would do (input / claims) |
| `actual` | What it verifiably returned (or the exception) |
| `verdict` | `grounded` \| `ungrounded` \| `silent_failure` \| `error` |
| `checks` | Score-shaped rows (`name`, `value`, `passed`, `source`) |
| `grounding_score` | Always `estimated: true` |
| `hash` / `prev_hash` | SHA-256 chain in-process |

Attached on the return value as `_agenko_receipt` (Python).

## Gate (`@approve`)

A **deterministic pause** before `when="irreversible"`. The SDK posts an approval request, polls the control plane, and either runs the function or raises `ApprovalError` / `ApprovalTimeoutError`. Non-irreversible `when` values pass through.

## Verdict

Not an LLM judge. Measured facts: empty output, claimed tool vs recorded tool, exceptions. Heuristics that are not hard facts stay estimated.

## What is not a concept (yet)

`why()`, harness objects, and eval CI are later milestones. Do not treat landing copy as the SDK contract — this site tracks what the code exports today.

## Agent instructions

This page: https://docs.agenko.app/core-concepts.md

Fetch `/llms.txt` for the full index. Search: `GET /api/search-docs?query=...`
