An AI agent engineer builds the systems that allow an AI agent to remember context, use tools, take actions and operate reliably in production. The role combines software engineering with model orchestration, data integration, evaluation, observability, security and cost control.
That platform layer became much more visible this week.
On 29 September 2026, OpenAI introduced Dots, persistent agents that can work across connected applications and continue projects without being directed through every step. Each dot operates on a cloud computer, carries context between channels and follows rules that determine which actions can proceed independently or require approval.
The job market shows what sits behind that experience. On 3 October, Bot Jobs contained 287 live vacancies, including 133 in Engineering.
The valuable skill is building the dependable system around the model.
Key takeaways
- AI agent engineering is a form of production systems engineering, not advanced prompt writing.
- Current roles combine backend development with orchestration, retrieval, memory, tool calling and evaluation.
- Reliability includes task success, latency, cost, permissions, traceability and recovery from failure.
- Python and TypeScript appear frequently, but strong architecture and debugging judgement matter more than one framework.
- A useful portfolio shows a complete working system, including failure paths and operational evidence.
What does an AI agent engineer do?
An AI agent engineer turns a model into a service that completes work safely and repeatedly. They design how it interprets a goal, selects tools, maintains state, retrieves information and decides when to continue, ask for help or stop.
The job may appear under several titles:
- AI agent engineer;
- agentic AI engineer;
- AI platform engineer;
- Conversational AI software engineer;
- applied AI engineer;
- LLM infrastructure engineer;
- backend engineer, AI platform.
Read the responsibilities rather than relying on the title. Memory, orchestration, tool calling, retrieval, model routing and AI observability usually signal the same career family.
Why do persistent agents need a platform layer?
A chatbot can receive a message, produce an answer and end the interaction. A persistent agent may work for hours or days, respond to changing information and use several external systems. That introduces engineering problems that a prompt cannot solve.
State and memory
The system must know what has happened, what remains unresolved and which information should persist. It also needs rules for expiry, correction and separation between users or projects.
Memory is not simply a transcript. Useful state may include verified facts, completed actions, pending decisions, tool results and the evidence behind them. Engineers need to decide what belongs in short-term context, durable storage or an authoritative business system.
Tools and integrations
Agents become useful when they can retrieve an account, update a ticket, schedule an appointment or run a workflow. Each tool needs authentication, an input contract, permissions and predictable error handling.
A production design must account for partial success. If an agent charges a card but fails before recording the order, repeating the entire action could make the problem worse. Idempotency, confirmation and recovery are user-experience requirements as well as engineering ones.
Reliability and observability
Traditional software can be tested against defined inputs and outputs. Agents add variable language, tool selection and model reasoning. Teams therefore need logs and traces that show what the agent received, which tools it chose, what evidence it used and why a task failed.
OpenAI’s 2 October guide to building with GPT-6 recommends measuring task success, latency and cost per successful task. It also covers context compaction, monitoring, data controls, long-running work and explicit boundaries for independent action.
The principle is platform-neutral: monitor the outcome as well as the model response.
Latency, throughput and cost
Every model call, retrieval step and external tool adds delay and expense. In voice AI, a few hundred milliseconds can change whether the exchange feels natural. At scale, unnecessary context or repeated calls can make an otherwise effective agent commercially unworkable.
Agent engineers choose models, cache stable context, run independent tasks in parallel and set budgets for time, tokens and tool use. The right design is rarely the largest model for every step.
Permissions and governance
An agent that can act needs an identity and defined authority. Engineers work with security, product and operational owners to decide which data it can read, what it can change and when a person must approve the action.
OpenAI’s Dots announcement describes custom rules, activity views and action review. Current vacancies also mention guardrails, governance, compliance and human oversight. These are becoming ordinary production requirements rather than separate policy exercises.
What are employers asking AI agent engineers to build?
The following vacancies were live when checked on 3 October 2026.
Senior Software Engineer, Conversational AI at Glia
Estonia, remote | Full-time
Glia is building a voice AI framework for banks and credit unions. The work covers conversation orchestration, tool calling, runtime architecture, grounding, guardrails, evaluation and end-to-end latency.
Evidence worth showing: a real-time agent with measured latency, documented model controls and a reproducible failure investigation.
Senior Software Engineer, AI Platform at Ridgeline
San Ramon, California | Full-time | US$153,000 to US$210,000
The role covers agent-builder capabilities, memory, cost guardrails, tool integrations, retrieval and Model Context Protocol connections. It requires production software experience rather than AI research credentials.
Evidence worth showing: an agent service with durable state, controlled tools, cost reporting and a clear interface other developers can reuse.
AI Engineer 4 at Capital One
Multiple US locations | Full-time | US$197,300 to US$245,600 depending on location
Capital One’s role spans LLM inference, agents, multi-agent workflows, similarity search, guardrails, evaluation, governance and observability. It also includes service-level objectives for latency, uptime and model-performance drift.
Evidence worth showing: a production scorecard that links technical measures with task outcomes, supported by architecture and incident decisions.
Agentic AI Senior Engineer at AstraZeneca
Cambridge, Massachusetts or Gaithersburg, Maryland | Full-time | US$137,349 to US$206,092
This role builds agents for biologics scientists and connects them with approved data, models and laboratory systems. AstraZeneca explicitly asks for authentication, logging, evaluation, error handling and human oversight. The listing closes on 14 October 2026.
Evidence worth showing: a domain-specific agent that uses governed data, explains its tool boundaries and records evidence another specialist can review.
Maven AGI’s current Senior Software Engineer, Platform Team vacancy adds telephony, resilient services and platform architecture to the picture. Together, these roles show that agent engineering sits at the intersection of backend systems, AI behaviour and operational ownership.
Which skills should an aspiring AI agent engineer learn?
Build the stack in layers:
- Software engineering: Python or TypeScript, testing, version control, APIs, databases and asynchronous work.
- Agent orchestration: state machines, tool calling, structured outputs, memory, retrieval and multi-step workflows.
- Production infrastructure: containers, cloud services, deployment pipelines, queues, caching and service-level objectives.
- Evaluation: representative tasks, expected actions, regression tests and outcome-based scoring.
- Observability: traces, logs, latency percentiles, token use, tool errors and cost per successful task.
- Security and governance: identity, least-privilege access, approval points, data boundaries, audit records and safe recovery.
- Domain understanding: enough knowledge of the user’s work to choose useful tools, recognise dangerous errors and measure the real outcome.
You do not need to master every vendor framework. Learn the recurring concepts and become comfortable moving between tools.
How do you build an AI agent engineering portfolio?
Build one narrow agent that continues across several steps. An appointment-management agent, account-support agent or research-monitoring agent is enough if the implementation is inspectable.
Create a portfolio pack with seven parts:
- Responsibility brief: Define the user, goal, authority and completion condition.
- Architecture diagram: Show the model, orchestrator, memory, tools, data sources and human route.
- State model: Explain what persists, what expires and how incorrect information is repaired.
- Tool contracts: Document authentication, fields, validation, side effects and retry behaviour.
- Reliability target: Set measures for task success, latency, availability, cost and incorrect action.
- Evaluation suite: Include routine cases, ambiguous instructions, unavailable tools, partial completion, hostile input and approval boundaries.
- Operations note: Describe monitoring, alerts, rollback, incident ownership and the first improvement you would test.
LangChain Academy’s free Introduction to LangGraph course currently offers 55 lessons and six hours of video covering state, memory, human intervention and deployment. Python Use it to build the working slice, then add your own operational decisions.
On your CV, describe the system and evidence rather than listing a framework:
- Built a persistent agent with structured state, approved tools and human approval for consequential actions.
- Defined service-level measures for task success, latency, cost and tool reliability.
- Added tracing and regression tests that isolated failures across retrieval, model and API layers.
- Designed recovery for timeouts, partial tool success and repeated requests.
Frequently asked questions about AI agent engineering careers
Do AI agent engineers need a computer science degree?
Not always, although many senior roles request a technical degree or equivalent experience. Employers care about production software evidence, architecture judgement and the ability to debug a complete system.
Which programming language is best for AI agent engineering?
Python is common for AI frameworks and evaluation. TypeScript is widely used for product services and tool integrations. Java, Go and other compiled languages appear in enterprise platforms. Choose one primary language and learn APIs, asynchronous execution and testing well.
Is an AI agent engineer the same as an ML engineer?
They overlap, but the emphasis differs. An ML engineer may focus on training, inference and model performance. An agent engineer usually focuses on orchestration, memory, tools, product behaviour and reliable operation around one or more models.
Can a backend engineer move into agentic AI?
Yes. Backend engineers already understand services, databases, APIs and failure handling. Add model behaviour, retrieval, evaluation and permission design, then build one end-to-end agent project.
What should an AI agent engineering portfolio prove?
It should prove that you can move beyond a demonstration. Show how the agent remembers, acts, fails, recovers, stays within authority and produces evidence that the task was completed correctly.
The agent is only as useful as the system around it
Persistent agents make an old engineering truth newly visible: capability without dependable operation is not a product.
The model matters, but so do state, integrations, latency, permissions, monitoring and recovery. Employers are looking for people who can connect those layers and take responsibility for what happens after the first successful demonstration.
Build one narrow agent all the way through. Then explore current Engineering roles on Bot Jobs, browse all Conversational AI opportunities and create a job-seeker profile so specialist employers can find you.
