What Matters
An AI agent developer is a systems engineer who happens to use models. Employers need people who can define tool contracts, manage state, protect credentials, evaluate outputs, observe failures and connect an agent to a real business process. Prompt-writing alone is not a durable career. A strong portfolio proves that the system works, refuses unsafe actions and can be reconstructed after an incident.
DN Evidence Block
Reviewed: 21 November 2026. Evidence base: Python, TypeScript, Rust, OpenAPI, MCP, OpenTelemetry, NIST AI RMF and production agent architectures.
Boundary: hiring requirements change quickly. The roadmap prioritizes portable engineering and operating skills rather than one vendor framework.
The Eight-Skill Stack
| Skill | What to learn | Portfolio evidence | Common weakness |
|---|---|---|---|
| 1. Python | APIs, data, async, testing and model SDKs | Typed service with tests and reproducible environment | Notebook that cannot deploy |
| 2. TypeScript/Node.js | Web APIs, streaming, auth and product integration | Production-style API and client | UI without backend controls |
| 3. Data and retrieval | Schemas, indexing, provenance, freshness and permissions | Source-linked answers with update logic | Vector search without evidence quality |
| 4. Tools and protocols | OpenAPI, MCP, webhooks and idempotency | Agent that calls bounded tools and handles errors | Tool access controlled only by prompt |
| 5. Evaluation | Test cases, scoring, adversarial inputs and regression | Eval suite with failure categories | Demo judged by one successful run |
| 6. Security | Secrets, least privilege, injection, sandboxing and audit | Threat model and permission tests | API key in model context |
| 7. Observability | Traces, metrics, logs, cost and incidents | Dashboard and reproducible failure trace | Only application print statements |
| 8. Domain workflow | Business rules, approvals, evidence and support | Real acceptance criteria and owner | Generic chatbot with no user outcome |
Language Choice
Python is the strongest default for model work, data and rapid experimentation. TypeScript is strong for APIs, browser products, streaming and ecosystem integration. Rust fits performance- and safety-critical components. Many production systems are polyglot.
Use DN’s Python vs Rust vs Node.js guide to choose by layer rather than ideology.
Learn APIs Before Agent Frameworks
Understand HTTP, JSON, authentication, retries, rate limits, webhooks, idempotency and schema validation. Agent frameworks change faster than these foundations. OpenAPI makes service contracts machine-readable; MCP standardizes a growing set of tool and resource interactions.
Build one tool server from scratch. Define input, output, errors, authorization and logs. Then connect it to an agent.
State, Memory and Retrieval
Distinguish conversation history, task state, long-term memory and authoritative records. Do not store everything in a vector database. Use a relational or document store for facts that need versioning and access control.
Retrieval quality depends on source identity, chunking, freshness, conflict handling and citations. An agent that retrieves the wrong document confidently is not improved by a larger model.
Evaluation Is a Core Engineering Discipline
Create test cases before the prompt stabilizes. Include expected outputs, acceptable alternatives, refusal cases, stale data, missing tools, prompt injection, malformed records and provider outage. Run regression tests after changing model, prompt, tool or data.
Measure task success, unsupported claims, tool error, latency, cost, human correction and incident severity. A single aggregate score hides important failures.
Security and Permission Boundaries
Keep secrets outside model context. Give each tool the minimum scope. Separate read, write, payment and administration. Use sandboxing, approved domains, rate limits and independent transaction policy.
Review the OWASP risks for model applications and apply NIST’s governance, mapping, measurement and management approach. A developer who can threat-model an agent is more valuable than one who only adds more tools.
DN Rule
The model should never be the final authority over its own permissions. Enforce identity, amount, counterparty, frequency and revocation outside the reasoning loop.
Observability and Incident Handling
OpenTelemetry organizes traces, metrics and logs as complementary signals. For agents, a trace should connect user request, model call, tool call, policy decision, output and outcome. Redact secrets before export.
Define alert thresholds, incident owner, rollback and customer communication. Keep enough evidence to reproduce the failure without retaining every sensitive payload indefinitely.
Four Portfolio Projects That Signal Real Ability
- Sourced research agent: retrieves official documents, cites each claim, handles conflicting sources and exposes freshness.
- Support triage agent: uses a controlled knowledge base, escalates uncertainty and measures correction.
- Paid API: OpenAPI contract, auth, metering, idempotency, billing and support status.
- Bounded financial simulator: proposes actions but uses external limits and a fake-money environment.
Publish architecture diagrams, threat model, tests, cost report, incident example and limitations. A polished frontend is optional; operating evidence is not.
A 90-Day Learning Roadmap
| Weeks | Focus | Deliverable |
|---|---|---|
| 1–2 | Python/TypeScript, HTTP, Git, tests | Small typed API with CI |
| 3–4 | Model API, structured output and retrieval | Sourced answer service |
| 5–6 | Tools, MCP/OpenAPI and state | Multi-step bounded workflow |
| 7–8 | Evals, injection and security | Regression and adversarial suite |
| 9–10 | Tracing, metrics, cost and deployment | Observable hosted service |
| 11–12 | Domain project and case study | Production-style portfolio project |
How to Apply for Roles
Show one deployed system and one serious postmortem. Explain trade-offs. Be able to discuss why a tool was denied, how a retry avoids duplicates, how an eval caught a regression and how data is deleted. Those answers distinguish an operator from a demo builder.
Look for roles in agent engineering, AI product, developer tools, model evaluation, applied AI, data engineering, security, observability and domain automation. Job titles will vary.
DN AI Agent Developer Readiness Test
Score your current portfolio.
Methodology and Limitations
DN derived the roadmap from current protocol, language, observability, risk and production-agent requirements. The test is not a hiring credential. Employers weigh experience, location, education, communication and domain fit differently.
Frequently Asked Questions
Do I need a computer-science degree?
Not always, but you need strong software foundations, evidence of shipped systems and the ability to reason about security and failure.
Which language should I learn first?
Python is the strongest first choice for model and data work; add TypeScript for product integration.
Do I need to learn every agent framework?
No. Learn APIs, state, evaluation, security and observability, then use frameworks selectively.
What portfolio project is best?
A bounded agent with real sources, tools, tests, traces and a clear user outcome is stronger than a generic chat demo.
What is the most overlooked skill?
Evaluation and incident handling. Production agents need proof that changes improve the workflow without creating new failures.
Primary Sources
- Python Documentation
- TypeScript Documentation
- The Rust Programming Language
- OpenAPI Specification
- Model Context Protocol
- OpenTelemetry Signals
- NIST AI Risk Management Framework
- OWASP Top 10 for LLM Applications
Editorial disclosure: no language, framework or training provider paid for this roadmap.




