Building the Blueprint for Agentic AI at the Edge
Author: Tina Tsou, LF Edge Board Chair
Why we’re doing this
Every enterprise rolling out AI agents at the edge is hitting the same wall: no shared way to connect agents to tools and data, no common runtime pattern across cloud, edge, and on-prem, and no consistent approach to auditability or incident response when something goes wrong. Add in the expanded attack surface that comes with agents talking to each other across environments, and it’s clear the industry needs a common reference, not another one-off stack.
That’s the gap LF Edge’s SuperAI SuperBlueprint is built to close: a joint effort between LF Edge and the Agentic AI Foundation (AAIF) to define an open, reusable reference architecture for agentic AI running on and against edge infrastructure.
Why LF Edge + AAIF
LF Edge already has a proven playbook for this kind of work. Projects like Akraino have shown that open, vendor-neutral blueprints can drive real adoption across use cases like 5G and AI. AAIF brings the agentic AI side of the equation, MCP, AGENTS.md, and goose, as neutral, open building blocks for interoperable agent behavior.
Putting these two communities together lets us tackle a problem neither could solve alone: as physical AI, real-time robotics, and distributed inference push more workloads to the edge, agents will need to move fluidly between edge and cloud resources, especially as power and cooling constraints push more compute out of centralized data centers.
What we’re building
The blueprint spans the full stack an agentic edge deployment actually needs:
- A multi-agent runtime with standardized MCP-based tool integration, supporting planner/executor/observer patterns and shared state across agents
- Cross-environment platform integration so workloads can move between cloud, edge, and on-prem without re-architecting
- Project-level conventions (AGENTS.md) extended for multi-agent coordination and lifecycle management
- Security and governance built in from the start: identity, audit logging, incident playbooks, and multi-agent trust boundaries
- Observability and AIOps, using OTEL as the backbone for everything from development to capacity planning
- A sandboxed execution layer for safely testing and validating agent behavior before it ships
- A unified agent orchestration layer: persistent, portable agent runtimes with secure agent-to-agent communication and elastic scaling across a distributed “Swarm” of cloud and edge nodes
Initial domains we’re targeting include industrial/OT edge, telco network ops, retail, physical AI, smart city and mobility, healthcare, and public safety, each a real-world proving ground for the same underlying architecture.
Where things stand
We wrapped Phase 1 with a draft charter, defined scope, and initial stakeholder alignment across LF Edge and AAIF. We’re now in Phase 2 (running through late August), building out the MVP reference implementation and first blueprint “stack,” alongside an evaluation harness for early benchmarks on real edge deployment targets.
Coming up later in this phase: a proof-of-concept session with participating partners to pressure-test agent orchestration, sandbox execution, and security controls end to end, plus a benchmark program measuring everything from orchestration latency to fault recovery and resource utilization across cloud, edge, and hybrid environments. Phase 3 will bring a hardened release and the start of partner pilots.
How to get involved
This is an open effort, and it’s still early enough that your input can shape real architectural decisions:
- Join the weekly technical calls, Thursdays, 6:30–7:00 PM Pacific, through October
- Contribute to the reference implementation as it takes shape this phase
- Bring a use case: if you’re running or planning agentic workloads at the edge, we want your requirements in the room
- Talk to us about pilot deployments: Phase 3 is looking for partners ready to test the blueprint against real workloads
Fragmentation is the default outcome when every team solves agent-to-edge integration on its own. SuperAI SuperBlueprint is our bet that a shared, open reference gets the whole ecosystem there faster, and we’d rather build it with more voices in the room than fewer.
Interested in contributing or piloting? Reach out through the LF Edge InfiniEdge AI project channels or join a Thursday technical call.
