Open-sourcing Yeda AI's Foundry agents: Yoca and Yopa
Two public Foundry agent examples show cited recommendations, human approval, deterministic fixtures, and governed action boundaries.
An agent drafts a typed, cited recommendation; a human reviews and signs off; a governed action writes to the system of record only after approval. Two slices of our Foundry agent work, just open-sourced on github.com/YedaAI, show that pattern end to end - the same one we use with customers. No background autopilot, no uncited claims, no silent writes.
The repos are the public, open-source slices - not toy mockups. Each one ships with a local OSDK fixture so you can clone, install, and click through the operator UI in under five minutes without a Foundry tenant or an API key. Pull a fresh model draft only when you want one, and only against your own key.
What's in the repos
Yoca_Palantir
Reviewer-Grade Compliance Evidence DeskYoca gathers evidence, cites policy, flags missing artifacts, and only publishes an attestation after a human reviewer signs off. It turns a 30-60 minute evidence-pack review into a guided sign-off flow with citations on every claim.
Highlights
- Cited findings: every claim ties back to an inspectable artifact citation.
- Typed actions: ApproveFinding and PublishAttestation gate the audit trail.
- Human-in-the-loop: no attestation is published if missing evidence is non-empty.
- AI agent surface fans out to extract and policy-lookup workers and assembles a typed Finding.
- Hallucination and injection guards with a deterministic fixture fallback when no API key is set.
Yopa_Palantir
Governed Order Exception AgentYopa reconciles conflicting order systems, proposes a cited next action, and requires human approval before any writeback. It collapses a 20-30 minute reconciliation task into a 60-90 second review flow.
Highlights
- Cited recommendations: the model proposes a next action with evidence pinned to each system of record.
- Governed writeback: nothing changes upstream until a human approves the action in the drawer.
- 8 policy-regression cases plus 20 golden cases gate every push in CI.
- AIP Logic, Functions, OSDK, Workshop, and Automate working together on one real flow.
- Runnable terminal example walks one exception through reconcile, recommend, and approve.
Why we open-sourced these
Most AI demos stop at a chat snippet. The interesting work, the work that decides whether a customer trusts an agent enough to wire it to a production system, happens after the prompt. It is in the citations, the typed actions, the eval gates, the human-approval drawers, and the deterministic fixture that lets a CI run pass without calling a model. We wanted that work to be readable end-to-end on real Palantir Foundry surfaces, so teams evaluating governed AI can see the pattern instead of guessing at it.
Both repos are Apache-2.0 licensed. The production versions, with tuned prompts, larger evals, tenant connectors, and deployment support, are what we ship to customers. If you want that, talk to us.