Intelligence Layer
Composable agent cards — prompt, skills, memory, MCP servers, and a typed tool library wired into one reusable, versioned object.
erdos_fai.intelligence
ErdosFAI · part of ErdosFL · Erdos Federated Agentic Intelligence
ErdosFAI builds, tests, reinforces, and governs AI agents — then connects them to the systems you already run and drives real work to follow-through. One observable, audited runtime: composable agent cards, multi-agent pipelines, human-in-the-loop safety, continual learning, real-time insights, and a connected execution layer.
How it works
Three steps to put agents to work across the tools you already run — without replacing any of them.
Wire in EHRs, labs, ERPs, scheduling, and messaging with 60+ pre-built connectors. Auth lives in your vault — connectors only declare capability and whether they can write.
erdos_fai.execution.connectors
Start from a healthcare- or operations-native template — or compose your own. Each declares triggers, risk thresholds, SLA & escalation ladders, and which actions require approval.
erdos_fai.execution.templates
Dry-run first, then go live: score risk, gate every write behind RBAC approval, route tasks, escalate on breach, and capture it all in a hash-chained audit trail.
erdos_fai.execution.routing
The agentic lifecycle
Compose agent cards from prompts, skills, memory, tools, and MCP servers.
BLUEScore agents against rigorous evaluation suites before they ship.
REDRun multi-agent pipelines with human-in-the-loop gates.
BLUELearn from every trajectory — agents that improve release over release.
GREENDrive work across connected systems with policy, cost, and a tamper-evident audit trail.
GREENEnterprise agentic technology
Each layer is a real Python package. The core is standard-library only, so it reads, runs, and ships anywhere.
Composable agent cards — prompt, skills, memory, MCP servers, and a typed tool library wired into one reusable, versioned object.
erdos_fai.intelligence
Multi-agent pipelines — chain agents into steps that pass artifacts forward, with approval gates between them. Low-code by construction.
erdos_fai.orchestration
Multi-approver human-in-the-loop governance, admin-configurable write gates, and PHI/PII-aware redaction on every input and output.
erdos_fai.safety
Trajectory-driven evaluation and optimization, dataset synthesis, and refined memories written back to agents — versioned upgrades.
erdos_fai.learning
Real-time cost & ROI dashboards, agent metrics, event-replay tracing, and a hash-chained, tamper-evident audit trail.
erdos_fai.insights
Connect real systems and drive governed workflows: 60+ connectors, workflow templates, risk scoring, task routing with SLA escalation, and RBAC — with dry-run simulation.
erdos_fai.execution
Don't hand your agentic future to a single model provider — own the lifecycle, end to end.
Integrations
Connect the systems you already run — across 13 categories — without ripping anything out. Every connector declares whether it can write, so the Safety Layer can gate side effects.
List the full catalog any time with erdos-fai connectors.
Use cases
Ship-ready templates that turn signals into governed follow-through. Browse them with
erdos-fai templates.
Watch vitals and labs for deterioration trends and escalate to rapid response before a code.
Surface discharge barriers and route prioritized remediation tasks to nurses and case managers.
Track critical lab and imaging results to closed-loop acknowledgement with audit.
Assess discharge readiness and coordinate post-discharge appointments and med reconciliation.
Surface expected capacity from admissions and discharges to plan transfers ahead of demand.
Optimize staff allocation against predicted census across ER, OR, and inpatient units.
Route cross-system purchase requests through policy-based, multi-level approvals.
Optimize utilization-review creation and prior-authorization for case managers.
How it fits together
Get started
# clone, then from the repo root
pip install -e ./erdos-fai
# live Claude calls (optional)
pip install -e "./erdos-fai[anthropic]"
from erdos_fai import Agent, AgentCard, Pipeline, Step
triage = Agent(AgentCard(
name="Triage",
system="You triage support tickets.",
))
pipe = Pipeline("support", [
Step("triage", triage, output_key="triage"),
])
print(pipe.run("Dashboard is down").final_output)
from erdos_fai import TEMPLATES, Simulator, Signal
wf = TEMPLATES.get("inpatient-deterioration")
# dry-run before anything touches a live system
report = Simulator().simulate(wf, signals=[
Signal("news2", 7, floor=0, ceiling=9),
])
print(report.render()) # executed=False, writes gated
No API key? It runs offline on the built-in echo provider. Try
erdos-fai demo and erdos-fai execute. Set
ANTHROPIC_API_KEY to run on claude-opus-4-8.
Deployment
ErdosFAI is open source and runs anywhere — the deployment tiers describe support and scale, not a paywall on the framework.
FreeApache-2.0
Self-hostsingle department
Talk to usmulti-site & ICU
The product suite
Compose deep agents — skills, tools, memory, MCP — with guardrails baked in.
Continuous scoring, regression detection, and per-metric reports.
Agents adapt in production, getting smarter with every run.
Connectors, templates, risk scoring, task routing, and SLA escalation.
HITL approvals, RBAC, write gates, cost control, and tamper-evident logs.
Build governed agents, connect them to your real tools, and put them to work safely.