organism-core · Python · AGPL-3.0
AI agents that earn
trust before they get it
Before an AI tool runs, organism-core settles one question: what does “done” actually mean here? After the action, it checks the result against exactly those criteria. Tools that repeatedly deliver get to do more on their own — tools that drift are demoted automatically.
Open source (AGPL-3.0), self-hostable — commercial license available.
How it thinks
Most agent frameworks ask “what should I do?” — organism-core first asks “how will I know it was done right?”
It researches the answer from six places before acting — the project’s own profile, lessons from past mistakes, related cases, the knowledge archive, domain patterns, and finally you. Only then does the work start. And when the result comes back, it gets scored against that list — every single time.
What sets it apart
Lessons that actually stick
When something goes wrong, the lesson is distilled, stored, and fed back in next time — even across completely different task types. Plenty of frameworks loop on feedback now; nobody else, as far as we can see, redistributes those lessons across domains as a built-in part.
Trust is earned, not configured
Every action type climbs five trust stages on its own track record — from "a human does this" to "runs autonomously". No global autonomy switch, no per-agent permissions file. Just demonstrated quality. (The research world calls this the Digital-Apprentice model — published as a concept in June 2026. This is a working implementation.)
And trust is taken back
If quality drops, the demotion is automatic — built into the architecture, not bolted on as policy. That revocability is exactly what the OWASP Agentic Top 10 (2026) asks for and almost nobody ships.
Done-criteria researched fresh, every time
Not hard-wired once in a config file — derived per action, per project, per moment, with the human as the final source when the system isn't sure.
The API in 30 seconds
Two methods. Everything else is brought by the harness.
from organism.adapter import BaseEffector
class GreetingEffector(BaseEffector):
name = "greeting_effector"
def define_done(self, request, context):
return {} # the DoD engine derives the criteria
def act(self, request):
return {"greeting_present": True}
A new action type starts at stage (b) proposed —
every action passes a human gate until the track record says otherwise.
The full runnable round trip lives in the README, guarded by CI.
Plays well with others
organism-core is the layer under your agent framework, not a replacement for it. LangGraph, CrewAI, Microsoft Agent Framework, self-evolving agents like Hermes — your reasoning agent runs as a tool inside the orchestration, and the orchestration wraps the gates, scores, and lessons around it. Approval buttons and rubric loops are everywhere in 2026 (the EU AI Act mandates human oversight anyway) — the part nobody ships is the whole path, fused: research what “done” means → gate → earn autonomy → validate → remember.
Born in production, not in a slide deck
organism-core grew out of a year of real production use in a working architecture practice — about 300 active projects, real mail, real invoices, real plans. Including the unglamorous part: a full taxonomy of how tool-to-tool hand-offs fail silently (no-ops that log nothing, schemas drifting apart, success signals that turn out to be test artifacts). The pattern that closes them — Receipted Transfer — is documented in the repo, failure taxonomy and all.
in evaluation
organism-core Cloud
The framework is open source (AGPL-3.0) and self-hostable. What we’re exploring on top: a hosted approval gate with audit reports, EU-hosted and GDPR-first, built for teams that need EU AI Act Art. 14 evidence without assembling it by hand — approval trails, autonomy stages, and provenance as exportable audit packs.
No dates, no pricing — we’re listening before we build.
Join the waitlist