Agentic AI System — Architecture (Skills & Tools)

High-level system diagram showing orchestrator, skills, tools, runtimes, data stores, and guardrails

Overview & Architecture

The design below is a general-purpose blueprint for an agentic AI system, drawn as a single reference diagram rather than a narrative report. It brings together several strands from the current agentic-AI literature: an orchestrator that interleaves reasoning with tool calls in the style of the ReAct pattern[1]; a skill/tool catalog that lets a model decide which external function to invoke and with what arguments, following the tool-use approach demonstrated by Toolformer[2]; and a vector database feeding retrieval into the model runtime, the retrieval-augmented generation pattern that combines a language model's parametric knowledge with an external, non-parametric index[3]. The event bus, sandboxed tool adapters, and audit log are this design's own choices for turning those patterns into a production-shaped system with policy enforcement and provenance tracking.

Architecture Diagram

Operator Console UI / Dashboard / Manual commands API / SDK Programmatic client & webhooks Agent Orchestrator Goal manager • planning • scheduler Planner / Policy Task Router Skills Catalog Skill metadata • signatures • constraints • WebFetcher • DataProcessor • Summariser • Planner Model Runtime Local LLMs • API-backed models • caching Skill Executor Executes skill code • isolates runs Event Bus / Task Queue Tool Sandbox & Policy Gate Web / HTTP Adapter Browser / Headless Filesystem / Shell External APIs & Tools External Services Third-party APIs • Web • SaaS • Internal DBs Observability & Monitoring Metrics • Traces • Logs Vector DB Embeddings • Retrieval State DB Agent state • sessions • checkpoints Audit Log Immutable actions • provenance Frontend / Operators Backend / Orchestrator Data stores Model / Cloud Runtime Security / Policy Message Bus / Queue
Diagram notes
Flow: Client → Orchestrator → Event Bus → Skill Executor / Model Runtime → Tool Adapters → External Services. Security & policy gates mediate tool access; observability and immutable audit capture provenance.

Component Overview & Subsystems

Skills
Discoverable units: metadata, I/O schema, security policy, sandbox requirements.
Tools
Adapters expose safe interfaces (web fetch, shell, file, APIs). All tool calls pass through policy & audit.
Runtimes
Model pool: local + API-backed with caching, routing, and cost controls.

References

  1. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. International Conference on Learning Representations (ICLR 2023). arXiv:2210.03629
  2. Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Zettlemoyer, L., Cancedda, N., & Scialom, T. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. arXiv:2302.04761
  3. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems (NeurIPS 2020). arXiv:2005.11401