Agent Skills Runtime is the reference implementation of ORCA (Open Cognitive Runtime Architecture). Its starting point is not to build another agent orchestrator, but to provide a runtime for "executable cognition": turning recurring agent reasoning logic into executable skills so that they are reusable, testable, observable, and portable across tools and model providers.
Purpose and Problem Background Most agent systems still encode critical logic in prompts and framework glue code, which brings recurring engineering pain points: reasoning logic is locked in prompt text rather than executable workflows; workflows are hard to reuse and even harder to test; contracts between steps are implicit and fragile; observability and auditability are often added after the fact; security and governance controls are inconsistent; and switching providers or frameworks often means extensive rewriting. ORCA's approach is to introduce a cognitive workflow execution layer.
Main Capabilities and Architecture
- Skills: reusable cognitive workflows declared as DAGs.
- Capabilities: stable, contract-driven interfaces with typed inputs and outputs, independent of the backend.
- Bindings: interchangeable execution backends that connect contracts to execution, including PythonCall, OpenAPI, MCP, and OpenRPC.
- Runtime: DAG scheduler + policy engine + cognitive state (CognitiveState) + execution trace (trace), making the reasoning structure explicit and executable while maintaining portability across backends. Repository terminology note: when referring to cognitive runtime terminology, COGIT is used to mean capability; however, in this repository's implementation, file paths, APIs, schemas, CLI, directories, and runtime internals still retain the canonical technical terms capability/capabilities and skill/skills to maintain architecture and compatibility.
Cognitive Classification The pure cognitive layer is intentionally narrower than the full runtime. The current classification distinguishes: pure cognitive capabilities (decision, evaluation, evidence, memory, perception, reasoning); compatibility interfaces (legacy or transitional names still present in the active registry during migration, such as eval.*); and operational capabilities (routing, delegation, workflow control, and other runtime helpers, which should not be counted as core cognition). The registry-level reference documentation for the classification is agent-skill-registry/docs/COGNITIVE_TAXONOMY.md, used to determine whether a certain capability belongs to the cognitive core or the operational layer.
Usage and Integration Local-first operation requires no API key, and deterministic Python baselines can be used for offline development and testing; the first run may take 30-60 seconds. The official demo skill is decision.make, showcasing a complete decision workflow under uncertainty: a multi-step pipeline (option generation, analysis, scoring, argumentation, validation), structured output (such as recommendation, tradeoffs, confidence_score, confidence_level, uncertainties, next_steps), risk-aware reasoning through explicit criteria and constraints, and support for trace-friendly execution aligned with ORCA observability goals. Note that the JSON output in the documentation is illustrative; actual output depends on the input context, bindings, and policy settings. Optional integration surfaces include: embedded SDK (lowest latency, in-process), HTTP API (service boundary, non-Python clients), MCP server (tool ecosystem and MCP hosts), framework adapters (LangChain, CrewAI, AutoGen, Semantic Kernel), native tool definitions (Anthropic, OpenAI, Gemini). When building skills, you can declaratively write skills as DAG workflows, reuse existing capability contracts, validate wiring and execution behavior, and package and contribute reusable workflows.
Advanced Features and Quality Gates Advanced features include: authentication and RBAC controls, Webhook events, plugin extension points, audit modes and runtime observability, CognitiveState v1 and cognitive hints, runtime-managed output envelopes (status, rationale, trace_ref), JSON Schema generation and validation, skill governance and consistency tools. The runtime's quality gate package targets pure cognitive capabilities, and can run a gate pack or only generate a scorecard; the main artifacts are artifacts/cognitive_e2e_contract_report.json, artifacts/cognitive_semantic_all_report.json, artifacts/cognitive_quality_scorecard.json, artifacts/cognitive_quality_gates_local_report.json. The documentation can be served locally.
Applicable Scenarios and Limitations Suitable for teams that want to decouple agent reasoning structure from prompts, need testable/observable/auditable execution, and want to maintain portability across frameworks, tools, and model providers; it can run independently, but its strongest use case is as a reusable execution layer beneath frameworks, tools, and model providers, rather than replacing all agent frameworks. The repository includes paper information: Beyond Prompting: Decoupling Cognition from Execution in LLM-based Agents through the ORCA Framework (Fernandez Alvarez, G. E., 2026), with DOI and SSRN links provided, and the paper page is docs/PAPER.md. The project welcomes contributions, see CONTRIBUTING.md; it uses the Apache 2.0 license, see LICENSE; for research citation methods, see software citation and CITATION.cff. The repository also includes a code of conduct and security policy. Specific fees, prices, platform coverage, and results are not provided in the materials.
