Append-Only Knowledge Graph Architecture for Multi-Agent Systems

X postMr. BuzzoniOpinion

Mr. Buzzoni (@polydao) outlines an append-only graph engineering architecture where dedicated agents extract and connect claims without updating existing rows. Using daily sealed log files, DuckDB, and human approval for merges, the system preserves complete auditability.

Key points

  1. Separated agent duties: extractor.md pulls claims from raw sources, linker.md proposes connections without merging, and checker.md evaluates contradictions.
  2. Single entry gate: write.py validates inputs and appends them to daily sealed JSONL files, while raw blobs are deduplicated via SHA-256 filenames.
  3. Strict append-only updates: In a database of over 12,000 claims, no record is modified; corrections are appended to let the log resolve mistakes.
  4. Derived indices: The DuckDB search index is completely transient and can be rebuilt from raw logs in eight seconds.
  5. Human-gated entity merges: Suggested node merges are held in merges.pending.jsonl and require explicit human sign-off.

Why it matters

When research swarms generate high-volume findings, overwriting existing records risks compounding extraction mistakes. A deterministic, append-only log ensures all connections trace back to raw documents and allows full reconstruction at any time.

Analysis

The architecture addresses key reliability issues in agent workflows by isolating the write path and enforcing immutable storage. Treating the graph database (DuckDB) as an expendable, derived artifact rather than the primary store provides fast query performance without sacrificing provenance or data integrity.

Original post

Mr. Buzzoni

@polydao · 30 August 2026

this is f*cking gold the cleanest graph engineering setup I've seen one agent extracts claims from raw sources, one proposes how they connect, one checks every claim against what's already there every claim gets appended, timestamped, sourced. nothing ever gets overwritten t.co/Gxua1WUvtD t.co/lDC4bR5A5R

606 likes18 replies1.1k bookmarks75.8k views

View on X
Selected replies (5)
  • @arcyton @polydao The step that decides whether a setup like that survives is re-extraction. Sources change, and if the graph cannot be rebuilt incrementally you end up with a snapshot that quietly ages out of usefulness. Does it handle updates, or is it a one-shot build?
  • @amaturefuturist @arcyton @polydao I see people in comments saying append. Yes? It will run into problems. The process is ok, needs enrichment step and cost work. Also capped a userbase of 1 - 1 domain and without shaped retreval it has a max usable size. Based off notes on a similar iteration ymmv
  • @itsthedonhashim @polydao @polydao wow, that's slick. makes data integrity a breeze. gonna try something like this for my next project.
  • @Argona0x @polydao gonna steal that never merges rule bro
  • @ArchiveExplorer @polydao ill go further - this isnt gold, its platinum thanks for sharing

Mentioned links

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • X post:Kirill

    From Loop Engineering to Graph Engineering: Deterministic AI Architectures

    Kirill describes an architectural shift in agent development from fragile single-agent loops to Graph Engineering. He argues that system reliability depends not on larger models or more agents, but on deterministic structures using nodes, edges, state verifiers, and knowledge graphs.

    27.3kbookmarks2.7mviews

    AI & LLMs· Opinion

  • X post:Mr. Buzzoni

    Claude Loop Engineering: Structured Agent Loops for Autonomous Overnight Runs

    Mr. Buzzoni (@polydao) outlines a file-based harness and directory architecture for Claude Code that enables AI agents to execute overnight tasks autonomously. Using clear contracts, verification rubrics, state tracking, and documented receipts, the setup is designed to run predictably without manual night-time interventions.

    373bookmarks29.7kviews

    AI & LLMs· Resource

  • X post:Mahax

    Graph Engineering: Why Agent Workflows Depend on Graphs

    Mahax cites Andrej Karpathy's claim that prompting is disappearing in favour of graphs, and provides a practical guide on graph engineering with Claude to optimise AI agent workflows.

    2.4kbookmarks240.8kviews

    AI & LLMs· Opinion

  • X post:codila

    Graph Engineering: Multi-Agent Systems and Dynamic Workflows

    Codila highlights a free two-hour Google course on graph engineering for multi-agent systems and outlines how to evolve beyond linear agent sequences into networked parallel graphs. Using Dynamic Workflows in Claude Code v2.1.154+, the guide demonstrates orchestrating hundreds of agents simultaneously without exhausting context windows.

    6.8kbookmarks465.4kviews

    AI & LLMs· Tutorial

  • X post:nicco

    Graph Engineering Over Prompting: Why Agent Workflows Require Graph Architectures

    A post by nicco highlights remarks by Andrej Karpathy asserting that prompting is an ephemeral interface and that graph architectures represent the true architectural layer for agent workflows. It pairs this with a practical guide by seeco on replacing sequential agent chains with explicit graphs.

    529bookmarks58.5kviews

    AI & LLMs· Resource

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.