Agentic Engineering Status¶
Current focus¶
Agent Skills, Agentic Loops and Agent Architecture are all complete at first-draft depth. The Agentic Engineering workstream is now complete as a first-pass learning map from capability design through runtime control to production architecture.
Current state¶
- AI Foundations is complete at first-draft depth.
- Agent Skills is complete with 12 detailed first-draft topics.
- Agentic Loops is complete with 12 detailed first-draft topics.
- Agent Architecture is complete with 12 detailed first-draft topics.
- Architecture coverage includes modular monolith, Hexagonal/Ports & Adapters, Clean Architecture, DDD/bounded contexts, context/state/memory boundaries, capability/tool architecture, RAG integration, security, queues, reliability, multi-agent trade-offs, evaluation and observability.
Decisions¶
- Follow the progression capability → loop → architecture.
- Treat agentic systems as normal software systems with probabilistic components inside explicit deterministic boundaries.
- The model proposes semantic decisions; application/runtime code owns canonical state, lifecycle, authorization, budgets, approvals, side-effect guarantees and hard completion rules.
- Keep context as a temporary purpose-specific projection rather than execution state or a database.
- Distinguish domain/operational state, run state, session state, long-term memory and model context.
- Treat tools as execution mechanisms behind application-owned capability/port contracts; keep provider SDKs and MCP/connectors in adapters.
- Organize modules around business capabilities rather than AI providers/frameworks.
- Prefer a modular monolith when it satisfies ownership, scaling, security and reliability requirements; extract services only for concrete boundaries.
- Prefer deterministic workflows with bounded agentic islands where overall control flow is known.
- Multi-agent architecture requires a real responsibility, permission, trust, scaling or ownership reason and should be evaluated against a simpler baseline.
- Use retrieval for reference knowledge and live tools/application services for current operational truth.
- Treat retrieved/tool content as data/evidence rather than trusted control-plane instruction.
- Long-running runs should use durable state, queues/event-driven resume, checkpointing, idempotency/reconciliation and explicit retry budgets.
- Evaluation and observability must cover the full trajectory: context, model calls, policies, capabilities, state transitions and final outcome.
Completed¶
- Agent Skills topics 1–12 — first detailed drafts complete.
- Agentic Loops topics 1–12 — first detailed drafts complete.
- Agent Architecture topics 1–12 — first detailed drafts complete.
Next steps¶
- Review/refine Agent Skills, Loops and Architecture as questions arise.
- Validate the patterns through implementation examples and architecture exercises.
- Choose the next separate AI workstream deliberately: RAG, MCP or Memory are the main candidates.