Everything is an atomic leaf
Work isn’t a tangled agent loop — it compiles to a DAG where every leaf is its own fault boundary: independently verifiable, parallelizable, graftable by best-of-N. The leaf is the atom of work here.
Open-source · a realm for crafted agents
Xihe is the platform. Every agent that dwells here is meticulously crafted, with a harness of its own. Reliability is engineered from the outside, not prompted into the weights. Wright is the first. More are on the way.
What Xihe is
A single developer, armed with the right scaffolding, can run a fleet. Xihe is the reference daemon for that idea: the orchestration, memory, skills, and guardrails that turn a raw model into a dependable collaborator, open-sourced as the substrate, so each agent that arrives can be a finished thing rather than a fresh start.
Work isn’t a tangled agent loop — it compiles to a DAG where every leaf is its own fault boundary: independently verifiable, parallelizable, graftable by best-of-N. The leaf is the atom of work here.
You don’t carry a pile of skills, commands, and workflows in your head — they’re all embedded in Xihe: the skills substrate, thinking modes, guardrails, and working loops, ready out of the box. The harness remembers how; you decide what.
Skills, genes, and learned facts live unified on one SQLite substrate. Every run leaves the substrate a little sharper than it found it. A flywheel, not a fixed prompt.
Built on Xihe v1 builds on a fork of pi — a lean agent runtime — and layers the harness on top: orchestration, memory, skills, guardrails. pi runs the loop; Xihe is what makes it dependable.
Three pillars
DAG-atomized leaves, best-of-N graft, a frozen prefix-cache that makes 256-way concurrency affordable.
A temporal knowledge graph (L0–L6), entropy-reducing curation, a skill-creator flywheel with provenance.
Falsifiable contracts, M1–M7 thinking modes, drift detection with auto-rewind.
The Wright
A Wright isn’t a prompt with a name. It’s a finished thing: its own harness, its own taste, its own guardrails, sharing only Xihe’s ground. We add them one at a time, and only when they’re ready.
decomposes intent into an atomic DAG
code, edits, tools, across any model
adversarial checks before anything ships
A marketing agent: campaigns, outreach, and the voice that carries your work outward.
A journey-planning agent: itineraries, logistics, options weighed and routed end to end.
A knowledge & database agent: ingest, organise, query, keep a knowledge base honest.