Principle AI Product Building Architecture Knowledge Systems

Files are the universal interface between humans and agents

Markdown and YAML files on disk beat databases because agents already know file operations and humans can inspect everything

@danshipper + @nicbstme — Agent-Native Architectures + Fintool · Feb 24, 2026 · 18 connections

Agents already know cat, grep, mv, mkdir. Users can inspect and edit what agents create. File paths are self-documenting — /projects/acme/notes/ beats SELECT * FROM notes WHERE project_id = 123. iCloud handles multi-device sync for free.

Nicolas Bustamante at Fintool stores user data (watchlists, portfolio, preferences, memories, skills) in S3 as YAML files instead of a traditional database. This makes everything debuggable with cat, versionable for free, and directly editable by users. The context.md pattern — a single markdown file as agent working memory — embodies this same idea and connects to why Context is the product, not the model.

This pattern is why Markdown skill files may replace expensive fine-tuning works in practice — when the knowledge layer is plain files, both humans and agents can read, write, and evolve it without deployment cycles. Part of why the file system wins for agents is that it is the state store bash already comes with. See Agents need a REPL that keeps state between steps — the harness decides whether a tool protocol works, not the protocol, which argues CLI beats MCP for that reason and not because of the protocol.

This extends beyond developer tooling into enterprise architecture: Agents eat your system of record — the rigid app was the constraint, not the schema — when agents replace rigid apps, files become the source of truth and databases become regenerable caches derived from those files.

Connected Insights

References (4)

→ Context is the product, not the model → Agents need a REPL that keeps state between steps — the harness decides whether a tool protocol works, not the protocol → Agents eat your system of record — the rigid app was the constraint, not the schema → Markdown skill files may replace expensive fine-tuning

Referenced by (14)

← Agents need a REPL that keeps state between steps — the harness decides whether a tool protocol works, not the protocol ← Context centralization is why coding AI works — git is a solved context repository, knowledge work has no equivalent ← Context is the product, not the model ← Markdown skill files may replace expensive fine-tuning ← Spec files are external memory that survives context resets ← Persistent agent memory preserves institutional knowledge that walks out the door with employees ← Similarity is not relevance — relevance requires reasoning ← Skill graphs enable progressive disclosure for complex domains ← Agents eat your system of record — the rigid app was the constraint, not the schema ← Context inefficiency compounds three penalties: cost, latency, and quality degradation ← When agents write the code, engineering moves into the harness — and documentation becomes the system of record ← Virtual filesystems replace sandboxes for agent navigation — intercept commands instead of provisioning infrastructure ← Knowledge systems need dual-layer storage — narrative depth and structured queries can't share a format ← An agent is its memory — keep its prompts, skills and tools portable and the agent can move between harnesses