Skip to main content

AI & Agent Workflows

AI agents produce retrieved sources, tool results, intermediate outputs, and final answers. Agent frameworks often store supporting context in private logs or application-specific databases, limiting independent verification and reuse across systems.

ALX Protocol represents agent artifacts as Blocks with deterministic identities and declared parent relationships. Other agents, evaluators, and reviewers can validate recorded dependencies without accessing the original orchestrator, provided the required Blocks are available.

The Problem​

Agent outputs often lack independently verifiable dependency records. Retrieval results, tool outputs, and intermediate steps remain in application-specific storage. Downstream agents and evaluators depend on the original orchestrator to reconstruct declared relationships between outputs and supporting context.

How ALX Applies​

Create Blocks for agent artifacts requiring deterministic identity or declared lineage:

  • Retrieved Sources: Metadata identifying documents, search results, or external resources.
  • Tool Results: Outputs from tools invoked during agent execution.
  • Intermediate Outputs: Results from earlier processing steps.
  • Final Answers: Outputs referencing direct dependencies through parentHashes.

Each Block declares direct dependencies. Earlier ancestors remain reachable through the derivation Graph. Root Blocks use an empty parent set.

Example: Agent Output​

The final answer Block contains application-specific content and declares the retrieved source and tool result as direct parents. The parent hashes are illustrative placeholders. A complete Block also contains protocolVersion, contentHash, and blockHash.

Block identity depends on content and declared parents. Identical content with different parent sets produces the same contentHash but different blockHash values. See Block Identity for canonicalization and hashing rules.

Content
{
  "type": "agent-output",
  "task": "Summarize the security review",
  "result": "Three high-priority findings."
}
Parent hashes
{
  "parentHashes": [
    "0x<source-block-hash>",
    "0x<tool-result-block-hash>"
  ]
}

Graph Validation and Attribution​

ALX provides two functions for examining declared agent dependencies:

  • validateGraph(seed, source, context) — Validates Block hashes and declared parent relationships. Returns valid, invalid, incomplete, or resource_limit_reached.
  • traceAttribution() — Returns a selected Block and reachable ancestors, including depths, path counts, and lineage completeness.

A valid Block can belong to an incomplete or invalid Graph. Missing parent Blocks indicate incomplete lineage. A missing parent Block can still exist in another system.

Protocol Boundaries​

ALX verifies deterministic identities and declared relationships. Applications remain responsible for additional requirements.

  • Retrieval and Storage: Applications manage retrieval, persistence, and indexing. XNDR Network provides separate discovery and indexing capabilities for observed ALX Blocks.
  • Execution Order: Parent relationships declare dependencies, not chronological execution order.
  • Content Validation: Block validation does not establish factual accuracy, safety, legality, quality, ownership, or approval.
  • Dependency Completeness: Graph validation cannot detect undeclared sources, tool calls, or other inputs.
  • Attribution and Economics: Lineage records declared relationships without assigning ownership, credit, compensation, or economic value.

Optional Block Attestations support Ed25519 or P-256 signatures over blockHash. See Signing for createSignedBlock() and validateSignedBlock().