JuankCadavid AI Agent Development — Decision Layer
This skill decides what to build an agent with and how to shape it. It operates nothing: framework APIs, SDK calls, and deployment mechanics belong to the skills and docs named in the deferral table below.
This file is the router. Read it, answer the three routing questions, then load only the references/ file(s) the task actually needs.
Scope & Composition
Owns: framework and runtime selection; agent architecture (state, tools, memory, human-in-the-loop points, multi-agent topology); the hosting choice for agent workloads; and the bridge from agent behavior to requirements.md, design.md, and tasks.md.
Does not own: API signatures, SDK usage, service configuration, deployment implementation.
Composition, not exclusion. This skill composes with the skills that own operation — it never displaces them. Their triggers legitimately overlap on agent-shaped tasks, so precedence goes by question type:
- Decision question — "which framework", "how should this agent be structured", "where should it run" → load this skill first.
- Operation question — "how do I call this API", "how do I configure this service" → load the runtime skill first.
- Both may be loaded in one task. A task that chooses and builds needs both, in that order.
Routing Heuristic — three questions
Answer these before opening the matrix; they shortlist the option in most cases.
- Where does it run? Your own process or container, a managed agent runtime, or inside an existing application. This is usually the hardest constraint, so it eliminates options first.
- How much orchestration control do you need? A prebuilt agent loop (tool-calling until done) versus an explicit graph you author (branching, parallelism, custom control flow). Control costs authoring effort — buy the loop when the behavior fits it.
- What must survive restarts, and where do humans intervene? Durable state across process death, resumable runs, and approval checkpoints are the requirements that most often force a framework instead of a plain SDK loop.
If the three answers do not converge on one option — or the task names a framework that must be justified against alternatives — open references/framework-selection.md.
Runtime Deferral
Runtime detail is out of scope by design. Availability is per-developer, not per-project: never assume a skill is installed, so every row names its substitute route.
| Question | Route |
|---|---|
| Bedrock / AgentCore operation, Harness, guardrails config | amazon-bedrock (when present; official AWS Bedrock docs otherwise) |
| Anthropic SDK, tool use, MCP | claude-api (when present; official Anthropic docs otherwise) |
| Lambda / ECS / API Gateway implementation of a chosen hosting | aws-serverless — packaged stack skill (when present; official AWS docs otherwise) |
| Framework API signatures, version-specific detail | Pinned official docs — see the ## Sources block of each reference |
references/aws-deployment.md owns only the choice between hostings. Implementing the chosen one is the third row's job.
References — load per task
| Read | When the task involves |
|---|---|
references/framework-selection.md |
Choosing or justifying a framework or runtime; the three routing questions did not converge; a user proposes an option you must place (including CrewAI or AutoGen) |
references/langgraph-patterns.md |
LangGraph already chosen or in use — state design, nodes and edges versus prebuilt agents, durable execution, HITL interrupts, streaming, memory, subgraphs |
references/akili-spec-mapping.md |
Writing spec documents for an agent feature — behaviors and evals as scenarios, architecture as design decisions, eval work as tasks with verification commands |
references/aws-deployment.md |
Deciding where an agent workload runs on AWS — AgentCore Harness versus Lambda versus ECS trade-offs for agent workloads |
More than one can apply: choosing a framework and then specifying the feature is framework-selection.md + akili-spec-mapping.md.
Maturity
The agent ecosystem churns faster than this skill is revalidated. Every factual claim here and in the references is pinned to an official-docs link, and the matrix carries a Last verified: date — check it before relying on a comparison.
Guidance not yet validated in a real build is marked inline with > Unvalidated: rather than omitted. The marker is greppable (grep -rn "> Unvalidated:" .claude/skills/ai-agent-development/), so the maturity boundary is auditable instead of implied. Treat marked content as a starting hypothesis to verify against the pinned source, not as settled practice.