The entire Multi-Agent ecosystem is built on sand: no lifecycle, no state model, no governance, no auditability, no reproducibility.
Every framework — LangGraph, AutoGen, MCP, OpenAI Swarm, Google Antigravity, CoreAgent — fails for the same structural reasons.
Today we’re releasing MPLP — Multi-Agent Lifecycle Protocol, the first open, vendor-neutral protocol that defines how agent systems must work: Context, Plan, Confirm, Trace, Role, Dialog, Collab, Extension, Core, Network — plus PSG, Drift Detection, Delta-Intent, AEL/VSL, and execution profiles.
This is not a framework. This is the TCP/IP moment for agentic AI.
Full spec, schemas, tests, and SDKs (Apache-2.0): https://github.com/Coregentis/MPLP-Protocol
Full rationale in first comment.
CoregentisAI•2mo ago
Hi HN,
Let’s stop pretending: The entire Multi-Agent ecosystem is built on sand.
Every vendor is hyping “Agent frameworks,” but no one has solved the fundamental problems that prevent agent systems from scaling beyond demos.
Today we’re releasing MPLP — Multi-Agent Lifecycle Protocol — the first protocol that directly attacks the unsolved problems the entire industry has been quietly avoiding.
This is not another framework. This is the TCP/IP moment for agentic AI.
The uncomfortable truth: Multi-Agent Systems today are fundamentally broken
Every major framework — LangGraph, AutoGen, MCP, OpenAI Swarm, Google Antigravity, Microsoft AutoDev, Meta AgentScope, CoreAgent — fails on the exact same six structural problems:
1. Hidden State = Hallucination time-bombs
LLM context ≠ State. No one has a deterministic state model. Not OpenAI. Not Google. Not anyone.
2. No lifecycle. No contract. No rules.
Agents “just call each other.” This works until step 8… then collapses.
3. Zero reproducibility
Same prompt, same model, same “graph”? Different results. No enterprise will trust this.
4. No governance layer
Nobody can answer: “Before this agent did X, who approved it?”
5. No audit trail
“Why did the system make this decision?” Today’s answer: “¯\(ツ)/¯”
6. No cross-system interoperability
A LangGraph agent cannot run in MCP. An MCP tool cannot run in AutoGen. Every vendor is building a walled garden.
Just like the pre-TCP/IP internet. Nothing talks to anything. Everything is proprietary and incompatible.
MPLP solves all six problems — not with a better framework, but with a protocol.
Frameworks compete. Protocols win.
Just like TCP/IP forced the world to agree on how networks work, MPLP defines how agent systems work.
What MPLP is (and why it’s fundamentally different)
MPLP is a vendor-neutral, lifecycle-complete, schema-first protocol that standardizes:
Context (the source-of-truth state model)
Plan (deterministic, inspectable execution plan)
Confirm (governance & approval gates)
Trace (immutable event history)
Role (agent capability definitions)
Dialog (structured communication)
Collab (multi-agent semantics)
Extension (tooling adapters)
Core (conflict resolution)
Network (transport semantics)
Plus:
Delta-Intent governance
Project Semantic Graph (PSG)
Drift detection
Verification Layer (VSL)
Action Execution Layer (AEL)
Structured Observability Events
SA / MAP Profiles (execution contracts)
No framework today has ANY of these. MPLP has ALL of them — as a stable, frozen v1.0 protocol.
MPLP is not competing with LangGraph / AutoGen / MCP — it’s exposing their missing layer LangGraph:
Great runtime graph. Zero governance. Zero lifecycle spec.
AutoGen:
Cute messaging abstractions. Not reproducible. No invariants.
OpenAI MCP:
Good tool-calling concept. Not a lifecycle. Not multi-agent. Not stateful.
Google Antigravity Agents:
Impressive demos. No formalism. No portability.
Anthropic CoreAgent:
Great theory. Closed. No schema. No lifecycle clarity.
Microsoft AutoDev / DevOps Agents:
Nice loops. Still prompt chains dressed as workflows.
Meta AgentScope:
Helpful scaffolding. Still framework-level. No lifecycle spec.
These are all runtimes without a constitution. MPLP is the constitution.
Frameworks will come and go, but protocols become the infrastructure of the world.
Why MPLP will likely become a standard (and why it’s inevitable)
If multi-agent systems are to become:
auditable
reliable
deterministic
repeatable
governable
safe
certifiable
enterprise-ready
cross-vendor
Then:
A lifecycle protocol is not optional — it is necessary.
There is no scenario where the future does not have an open protocol governing agent execution.
If the community adopts MPLP, it will become the TCP/IP of the Agent Era — the layer every agent framework eventually must speak.
Why we released MPLP as open source (Apache 2.0)
Because a protocol can only become a global standard if:
no company owns it
anyone can implement it
it is stable and frozen
schemas are public
invariants are testable
contributions follow a governance process
This release includes:
Full protocol spec
All schemas
Normative test suite
Multi-lang SDKs
Frozen invariants
Governance model
Compliance guidelines
This is not a paper. This is a full working protocol.
Links
GitHub: https://github.com/Coregentis/MPLP-Protocol
Docs: https://coregentis.github.io/MPLP-Protocol/
Spec Overview: /docs/00-index/mplp-v1.0-protocol-overview.md Schemas: /schemas/v2/ Quick Start: /docs/08-guides/mplp-quickstart-5min.md
What we want from HN
Tell us:
What breaks your current multi-agent system?
Where do your agents drift?
Which failure modes should MPLP mandate?
Which frameworks should we build adapters for first?
What must v1.1 include to become unavoidably standard?
We’re here to listen. Rip this apart. Make it better.