Stop building knowledge agents like coding agents: a 6-layer architecture for grounded knowledge work
MaryamMiradi · x · 2026-09-19
Researcher Maryam Miradi argues most AI agent architectures are overly shaped by coding agents, but code is a special knowledge type with filenames, functions, and explicit dependencies — while legal, medical, financial, or research agents start from an intent and must discover what information matters.
Her design approach for knowledge agents includes:
- Decompose the intent: turn open-ended requests into explicit research questions; identify entities, constraints, time periods, jurisdictions, and missing facts — don't blindly search the original prompt
- Plan the search: decide which evidence is needed and where to find it
She proposes a 6-layer architecture for grounded knowledge work, distinct from coding-agent workflows.
Related event: Knowledge Agents Need Their Own Architecture, Researcher Argues(2 posts)→
More from coding & agent
- Dev shares funny moment: coding agent Astra wrote the whole backend in Python inside TanStack Start — kevinkern · 2026-09-19
- BuilderIO's skills repo (4.3k stars) adds composable agent skills via one npx command — tom_doerr · 2026-09-19
- Seroter's reading list: Zed's Delta kills PRs, agents pick the infra — rseroter · 2026-09-19
- Antigravity 2.15.0 ships custom-agent tweaks: toggle default prompts, persist agent choice — gaganghotra_ · 2026-09-19
- Tsinghua's C2C Lets LLMs Skip Text and Merge KV-Caches Directly, 2.5x Faster with +14.2% Accuracy — anselm · 2026-09-19
- Apple publishes iPhone Duo development docs alongside Xcode 27.1 beta — JordanMorgan10 · 2026-09-19