Open operational infrastructure for production AI systems.

The AI Operations Specification comes before SDK ergonomics, package features, dashboards, or integrations.

Ecosystem roadmap

DeepAgentLabs is built specification-first: define the language of AI operations before implementing package-specific behavior. The AI Operations Specification is the foundation; the ecosystem packages are reference implementations of that foundation.

It gives multiple tools a shared operational model so they can represent the same run consistently. The specification stays above any one package, allowing third parties to implement the contract without depending on the Python packages directly.

flowchart TB
    Spec["AI Operations Specification"]
    Spec --> Concepts["Core Concepts"]
    Spec --> Conventions["Semantic Conventions"]
    Spec --> Schemas["JSON Schemas"]
    Spec --> Versioning["Versioning"]
    Spec --> Examples["Examples"]
    Spec --> Extensions["Extension Model"]
    Spec --> Implementations["Reference Implementations"]
    Implementations --> Lens["AgenticLens"]
    Implementations --> Chaos["Agentic Chaos"]
    Implementations --> Sidecar["Agentic Sidecar"]
    Implementations --> MCP["DeepAgent MCP"]
    Implementations --> Tower["AgenticOps Control Tower"]

The roadmap's initial operational objects:

The Phase 1 questions also name an LLM call, tool call, memory operation, and RAG retrieval.

At Phase 1, the focus is on definitions, relationships, and terminology, not Python APIs or exporters.

The specification defines a standard; each component has a distinct role against that shared model.

DeepAgentLabs is building an open, modular ecosystem for developing, operating, evaluating, and governing AI agents across real-world environments.

The ecosystem brings together a unified Control Tower, shared AI Operations Specification, and specialized open-source components for observing, evaluating, supervising, testing, and connecting AI agents.

flowchart TB
    Spec["AI Operations Specification"]
    Spec -->|instruments and exports| Lens["AgenticLens"]
    Spec -->|extends with resilience evidence| Chaos["Agentic Chaos"]
    Spec -->|governs decisions against| Sidecar["Agentic Sidecar"]
    Spec -->|exposes through MCP| MCP["DeepAgent MCP"]
    Tower["AgenticOps Control Tower"] -. "operator-facing control plane above ecosystem capabilities" .-> Capabilities["Ecosystem capabilities"]

The Control Tower does not replace Lens, Chaos, Sidecar, or MCP.

• Runtime Agnostic — works across different agent frameworks and runtimes
• Framework Agnostic — integrates with the tools and frameworks teams already use
• Modular & Open Source — independently usable components with open-source foundations
• Human + AI Operable — designed for real-world teams and human oversight

Together, these components provide an end-to-end approach to observe → evaluate → supervise → test → connect AI agents.

The phase titles and sequence below follow the roadmap. The phase list describes goals, not completion status; status is called out separately only where the roadmap states it.

CURRENT FRONTIER · Agentic Chaos

Structured experiment traces/reports covering hypothesis, injection point, fault, observed behavior, recovery outcome, verdict, and provenance; plus synthetic test scenarios for prebuilt known-bad agent behaviors.

This is the roadmap's explicitly named current frontier in its recommended build order. The consolidated priorities also list Agentic Chaos resilience benchmark fixtures/datasets and parallel/sharded test execution under Implement Next.

The MCP server supports local stdio through the deep-agentic-core-mcp PyPI package and hosted, authenticated HTTP at https://mcp.deepagentlabs.io/mcp. Get a user API key at mcp.deepagentlabs.io. Version 0.3.0 adds the optional http installation extra, user-scoped workflow state, signup/key lifecycle, and Sidecar discovery. Hosted HTTP disables chaos script execution. Package publication and AWS deployment share a version-tag release workflow; AWS deploys only after PyPI publication succeeds.

These labels preserve the roadmap's own status terminology.

1. deep-agentic-core-mcp — sessions, diagnostics, tool metadata, prompts, verification · shipped in 0.2.0
2. agenticlens — provenance/evidence, next-step recommendations, OTel, layer enforcement, conformance CLI · shipped in 0.4.0
3. agentic-chaos — structured reports/traces, synthetic scenarios · current frontier
4. ai-operations-spec — provenance/evidence/report semantics, naming rules, conformance requirements and fixtures

The long-term goal is not just a Python toolkit. It is an open operational standard for AI systems with a shared object model, shared semantic conventions, shared schemas, shared examples, additive extensions, and multiple interoperable implementations.