Platonic Core
Core Rust harness primitives for disciplined, replayable agent execution.
Platonic Core is not a chatbot, personal assistant, workflow SaaS, or “Agent OS.” It is the small kernel underneath those things: typed context assembly, tool-call boundaries, policy decisions, event logs, and replay/audit surfaces.
Referential.ai is currently the public site and showcase for Platonic, Platonic Core, and Plato Agent; any future hosted or commercial role remains undecided.
Naming
The workspace naming authority owns the hierarchy and exact forms. Platonic Core is the framework's sans-IO kernel.
Design stance
- Every side effect is typed.
- Every context byte has a lane and a budget.
- Every run is an event log first; transcript is a derived view.
- Provider fallback is a policy event, not a silent retry.
- Tool output is structured data plus a short summary; raw logs become artifacts.
- Memory is scoped, typed, source-backed, and budgeted — not a junk drawer.
- The core stays small. Gateways, dashboards, cron, skills, voice, and platform adapters belong outside the kernel.
Current crate contents
The crate is intentionally modular inside the kernel:
ids identifier newtypes
message model-facing message primitives
context lane-labeled context packs with budget validation
policy effect classes and policy decisions
tool tool-call and tool-result boundaries
event durable harness event ledger
run pure run state machine
projection pure run readback projection
error shared error types
The public contract is documented in the crate rustdoc.
Minimal host
Run the public kernel through one approval-gated tool turn and replay the resulting ledger:
The example simulates model and tool IO. It performs no file write or network call.
Verify
Status
Seed kernel. The run contract is implemented as a pure multi-turn state machine with replay-validated readback projections; outer apps still own IO, providers, tools, stores, gateways, renderers, and schedulers.
License
Licensed under either of
at your option.