DeepAgentLabs

Building in the open

AI operations need a shared language.

DeepAgentLabs is a specification-first ecosystem for understanding how AI systems run, what they cost, and how they fail.

Explore the specification See the ecosystem
Language-neutralFramework-agnosticLocal-first
ECOSYSTEM / 01LIVE
00
AI Operations SpecificationShared operational contract
01AgenticLensObserve + optimize
02Agentic ChaosStress + validate
03Agentic SidecarSupervise + guide
04DeepAgent Control TowerCoordinate + govern
05DeepAgent MCPExpose + operate
One modelMultiple implementations

The premise

Most AI tooling starts with a dashboard. We start with the operational contract underneath it—so every runtime, evaluator, and resilience tool can describe the same system consistently.

Reference architecture

The ecosystem, as one operating model.

The diagram below is the clearest representation of how users, runtimes, the control plane, the specification, and the packages fit together.

DeepAgentLabs reference architecture showing users, console, DeepAgent Control Tower, AI Operations Specification, Agentic Lens, Agentic Sidecar, Agentic Chaos, MCP Server, exports, and platform principles.

The ecosystem

Independent tools.
A common foundation.

Six connected building blocks, each useful on its own and more powerful when combined through the shared specification.

FOUNDATIONv0.1.0 · draft

AI Operations
Specification

The language-neutral model for workflows, runs, agents, model interactions, evaluations, safety, reliability, and incidents.

Read the standard
OBSERVABILITYv0.4.0 · Python

AgenticLens

Profile, evaluate, and compare AI workflows locally. Token usage, latency, cost, evaluation gates, and evidence-backed opportunities to reduce waste.

View AgenticLens
RESILIENCEv0.4.0 · Python

Agentic Chaos

Inject provider failures, corrupted outputs, tool errors, and agent faults before production does.

View Agentic Chaos
SUPERVISIONv0.2.0 · Python

Agentic Sidecar

Companion intelligence for autonomous agents with real-time supervision, governance, escalation, and human approval loops.

View Agentic Sidecar
CONTROL PLANEv0.2.0 · Python

DeepAgent
Control Tower

The operating console and control plane for agent registry, capability discovery, configuration, and a unified control API.

View Control Tower
CONTROL PLANEv0.2.0 · MCP

DeepAgent MCP

A unified MCP interface over the shared model for AgenticLens, Agentic Chaos, and Agentic Sidecar workflows.

View the MCP server

Specification first

Define meaning before implementation.

Workflows and runs are not the same thing. A retry is not automatically an incident. A completed run can still fail an evaluation. The specification makes these boundaries portable.

Review the v0.1 concepts
  1. 01
    Shared object model

    Consistent concepts for runtime activity and operational evidence.

  2. 02
    Semantic conventions

    Stable AI-native event meanings, independent of framework or transport.

  3. 03
    Portable artifacts

    Schema-backed evidence any conforming tool can produce or consume.

  4. 04
    Additive extensions

    Room for new systems without fragmenting the common core.

Open roadmap

From vocabulary
to interoperability.

Full roadmap
v0.1Core conceptsNow
v0.2RelationshipsNext
v0.3SemanticsPlanned
v0.4JSON SchemasPlanned
v0.5VersioningPlanned
v1.0Stable standardNorth star

Build with us

Help define how AI systems are operated.

The standard is early by design. This is the moment for framework authors, AI engineers, and reliability teams to shape it.

Join the discussion Follow on GitHub