Palantir's agent architecture docs decoded: six design decisions for a reliable agent harness
blaizedsouza · x · 2026-09-27
Drawing on Palantir's public architecture docs, the author distills six decisions for turning a model into a worker you can leave alone — most of them copyable into your own stack:
- Ontology over raw data: agents see business objects (orders, customers, assets) and a fixed list of allowed actions — typed tools like updateorderstatus, never raw SQL.
- A layer between agent and model: mask PII before prompts leave, cache responses, retry failures, track rate limits and token usage; route all calls through one gateway (LiteLLM or a custom wrapper).
- Swappable models: a model catalog plus bring-your-own-model, with the model name in config so switching is a one-line change.
- Three trigger modes: schedule, event, and (truncated) human entry points.
The linked article "How to Design an Agent Harness" targets familiar failure modes: agents claiming the job is done without running tests, forgetting rules after twenty minutes, and retyping the same instructions every session.
More from coding & agent
- AI now leads a quarter of tasks at Anthropic, up from 1% in February — victor_explore · 2026-09-27
- $800 litter-picking robot MOSS hits V0.3, open-source V0.4 already printing — AIFlow_ML · 2026-09-27
- Matt Shumer buys a 3D printer for Opus 5.5, teases wild demos — mattshumer_ · 2026-09-27
- Xiaomi open-sources RL environments on Hugging Face, potentially worth millions — burny_tech · 2026-09-27
- Antigravity 2.0 adds dedicated /plan mode with human-approved implementation plans — burny_tech · 2026-09-27
- Developer runs 39 terminals with 30 AI subscriptions in parallel agent spree — Scobleizer · 2026-09-27