Agentic AI Developer

WHAT THIS ROLE ACTUALLY IS

You build autonomous AI agents and multi-agent systems that execute complex workflows without human intervention between steps. Not chatbots. Not single-turn LLM wrappers. Agents — systems that plan, reason, take actions, use tools, handle failures, and complete multi-step tasks in production environments that enterprise clients and government agencies depend on. You understand that an agent that works in a demo breaks differently than one that breaks in production at 2am. You build for the failure case first. You define what the agent does when a tool call fails, when the LLM returns an unexpected output, when a dependency is unavailable. You do not leave those decisions to the runtime.

WHAT YOU DO EVERY SINGLE DAY

• Read every written specification before building — every agent has a documented scope, a defined tool set, a failure handling protocol, and acceptance criteria before you write the first line of code
• Build multi-agent systems using LangChain, AutoGen, CrewAI, or equivalent — you have used at least one in a production deployment and you can explain what broke and how you fixed it
• Design and implement agent tool integrations with error handling, retry logic, and logging on every single integration without exception
• Build agent orchestration layers — task decomposition, agent routing, inter-agent communication, state management, and output validation
• Implement agent memory systems — short-term context management, long-term vector storage, episodic memory for multi-session workflows — optimized for token cost and retrieval performance
• Build human-in-the-loop checkpoints where specifications require them — government and enterprise deployments require human approval gates before consequential actions
• Implement full observability on every agent — every reasoning step logged, every tool call recorded, every decision traceable. An agent without an audit trail does not deploy to a government environment
• Write tests for agent behavior — deterministic tests where possible, evaluation harness testing where outputs are non-deterministic. Test failure paths explicitly
• Document every agent — architecture, tool inventory, failure modes, escalation logic, and deployment requirements
• Submit a written daily standup update every working day without exception

TECHNICAL REQUIREMENTS — CORRECTED FOR GRADE LEVEL I–II (4–6 YEARS MINIMUM)

Agent Frameworks (3 yrs total / 2 yrs production): Minimum 3 years of professional experience building AI agents using LangChain, AutoGen, CrewAI, or equivalent — of which at least 2 years must be in production deployments serving real users with real uptime requirements. You have extended frameworks, worked around their limitations, and built custom components where they fell short. You can explain the architectural tradeoffs between ReAct, plan-and-execute, and multi-agent patterns from production experience — not from reading papers.

LLM Integration (4 yrs AI systems / 2 yrs direct LLM production): Minimum 4 years of professional experience integrating AI systems into production — including at least 2 years of direct LLM API integration. You have managed production LLM costs, handled context window constraints under real usage patterns, built retry and fallback logic tested by actual production failures, and dealt with model deprecation cycles in a live system.

Python (5 yrs professional production): Minimum 5 years of professional production Python — your primary language. Typed, tested, documented, async where appropriate, structured for long-term maintainability. You have written Python that senior engineers reviewed and found nothing to correct. You have also reviewed Python written by others and found things to correct — and you documented them specifically.

Observability and Logging (3 yrs total / 1 yr agentic-specific): Minimum 3 years of implementing observability systems for AI or software in production — of which at least 1 year must be specific to agentic AI observability. You have used observability data to diagnose a production issue and you can describe the specific incident, what the logs showed, and what you changed as a result.

Security in Agentic Systems (2 yrs production): Minimum 2 years of implementing security controls in AI systems deployed to enterprise or government clients — prompt injection defense, agent sandboxing, tool permission restriction, CUI data handling in automated pipelines. You have thought adversarially about your own systems and you have identified attack vectors and built mitigations before deployment — not after an incident.

Multi-Agent Architecture (2 yrs production): Minimum 2 years of designing and deploying multi-agent systems in production — supervisor-worker architectures, parallel agent execution, and consensus-based decision systems. You have seen multi-agent failure modes in production — communication overhead, state synchronization errors, compounding errors across agent hops — and you have built recovery mechanisms.

Memory Systems (2 yrs production): Minimum 2 years of production experience implementing agent memory systems — short-term context management under token budget constraints, long-term vector storage with real document corpora, episodic memory across sessions. You know your retrieval latency, your embedding cost per document, and your recall accuracy at the p95 level for at least one production system.

WHAT WE REQUIRE — NO EXCEPTIONS

• 4–6 years of AI and software development experience with at least 2 years specifically in agentic AI systems deployed to production
• You have built and deployed at least one multi-step AI agent serving real users with real uptime requirements — you can describe how it handles every failure mode
• You have implemented agent observability — every reasoning step and tool call is logged and traceable in your production systems