All roadmaps

AI Agents

From chatbots to systems that get work done.

Tool use, agent loops, memory, orchestration frameworks and MCP, multi-agent patterns, and the evaluation and safety discipline that keeps autonomy shippable.

Your progress0/34 topics · 0%

Progress is saved in this browser.

5 stages34 topics₹0 to follow
Skill certification coming soon

Filled nodes are the focus for this track; outlined ones can wait. Tap a topic to open it and mark it done, learning or skipped.

AI Agents in 2026Related roadmapsAI Engineer roadmapPrompt Engineering roadmapAPI Design roadmapProve it as you goReading is half of it. The free, scoredAI Engineer certification covers theground this roadmap leads to, and the AIEngineer job simulation puts it to workon real tickets - evidence a recruitercan check. Mark topics done, learning orskipped as you go; your progress carriesinto the classic view.AI Engineer certificationJob simulationClassic viewKnow the demand before thesyntaxAI Agents in 2026What to learn in 2026Tool use andloopsGuardrails andevalsMCP and protocolsWhich roles use itAI engineer andproductOps andautomationFresher vs switcherWhat companies expectThe hiring barWhat interviewstestProof they verifyWhat gets you hiredA 90-day planProjects thatget readCertification and proofWhat agents are - and when notto build one.Agent FundamentalsWhat makes an agentWorkflows vs agentsAgent use casesThe mechanics of action.Tools & LoopsTool / function callingThe agent loopReAct & planning patternsMCPLong tasks, multiple actors.Memory & OrchestrationMemory & stateMulti-agent patternsHuman-in-the-loopEvaluation and guardrails.Ship Agents SafelyEvaluating agentsSafety & sandboxingObservabilityFresher or experienced?Both tracks hold every topic. On the Fresher track thefoundations are filled and the later stages are outlined - dothem when a project needs them. On the Experienced track thefoundations are outlined and the later stages are filled. The2026 section at the top applies to both.Prefer a list?The classic view shows the same 4 stages as a sequentialchecklist with time estimates. Your marks are shared betweenthe two views.Open the classic view

Everything on this roadmap

Every stage and topic on the map, in reading order. Click any topic to open it.

1.Know the demand before the syntax

2.What agents are - and when not to build one.

3.The mechanics of action.

4.Long tasks, multiple actors.

5.Evaluation and guardrails.

The AI Agents roadmap, topic by topic

Every topic on the map in reading order, with what to learn and the resources we recommend. Main topics are open; open any subtopic to read it. Select a topic name to open it in the topic panel, where you can track your progress.

  1. 1. Know the demand before the syntax

    AI Agents in 2026

    2026

    AI agents - models that plan, call tools and act over several steps - are the fastest-growing capability in software and the one most likely to be done badly. Gartner predicts a third of enterprise applications will include agentic AI by 2028, up from under one percent in 2024, and that at least 15 percent of day-to-day work decisions will be made autonomously by then; it also predicts more than 40 percent of agentic AI projects will be cancelled by the end of 2027 over cost and unclear value. Lightcast counted roughly 90,000 US job postings mentioning agentic AI skills in 2025, up more than 280 percent, at a salary premium over AI and IT postings generally. The four nodes answer what to learn, which roles use it, what companies expect and what gets you hired.

    What to learn in 2026

    Learn how an agent loop works - tool use, planning, state - then the guardrails and evaluations that make one safe to run, and then the protocols, above all the Model Context Protocol, that connect agents to tools and data. The sub-nodes order the three.

    Tool use and loops

    An agent is a loop: the model reads the goal and the state, chooses a tool, the tool runs, the result goes back into the context, and the loop continues until the job is done or a limit is hit. Learn to build that loop by hand with a model API before any framework - tool definitions, structured calls, state and memory, planning and reflection, stopping conditions - so that you can debug what the framework hides. Gartner's forecast of agentic AI in a third of enterprise applications by 2028 is demand for people who understand the loop, not just the demo.

    Guardrails and evals

    Guardrails are what let an agent run unattended: bounded tool permissions, idempotent actions, checkpoints and human approval for anything that sends, pays or deletes, budgets on steps and spend, timeouts, logs of every call, and evaluation sets that replay real tasks and score the outcome. Gartner's prediction that more than 40 percent of agentic projects will be cancelled by 2027 is about agents that cost more than they returned or could not be trusted; Postman's 2025 report found half of developers naming rogue agent calls their top API security concern. The engineer who ships guardrails first is the one whose agent survives.

    MCP and protocols

    AI

    Agents reach the world through protocols: the Model Context Protocol has become the common way to expose tools, data and prompts to agents from any provider, with transports, authorization and security concerns of its own; function-calling schemas, tool descriptions and structured outputs are its vocabulary. Postman's 2025 report found 24 percent of developers already designing APIs with agents as consumers. Learn to build an MCP server for a real system, to describe tools precisely, and to secure them - because an agent with a badly described tool will use it badly.

    Which roles use it

    AI engineers and AI product managers build agents into products; operations and automation specialists deploy them inside workflows. The sub-nodes name both groups and lay out the fresher and switcher routes.

    AI engineer and product

    AI engineers build agents into products and AI product managers decide what they should do; both are hiring fast - LinkedIn's Jobs on the Rise 2026 ranked AI engineer the fastest-growing US title with postings up 143 percent, and Lightcast found agentic AI skills in roughly 90,000 US postings in 2025 at a premium over general AI roles. These roles want the loop, the guardrails and the evaluation built properly, inside a product with real users. The AI Engineer roadmap is the engineering path; the Prompt Engineering roadmap is the context half.

    Ops and automation

    Operations, support, finance, marketing and IT teams deploy agents inside their workflows - triage, drafting, reconciliation, research, scheduling - often through automation platforms and agent builders rather than code. Microsoft's 2025 Work Trend Index found 46 percent of leaders already using agents to automate workflows and describes a new role of agent boss for every employee. The skill in these roles is scoping an agent to a job that pays off, bounding it with approvals, measuring it and knowing when to turn it off.

    Fresher vs switcher

    A fresher should build one bounded agent that does a real job for a real person, with approvals, logs and an evaluation, rather than a general assistant; HackerRank's 2025 report found early-career hiring the hardest hit, and a working, measured agent is a rare proof. A switcher from backend engineering has the systems half and should learn the model half; a switcher from operations has the workflow and should learn the loop. On this map freshers get the basics filled and the protocol stages outlined; experienced learners get the reverse.

    What companies expect

    Companies expect an agent that does a bounded job reliably, with tools it is allowed to use, approval steps where actions matter, logs of everything it did, an evaluation that proves it works and a cost that pays off. The sub-nodes cover the bar, the interview and what gets verified.

    The hiring bar

    The bar is an agent you would let run while you sleep: a bounded job, tools scoped to it, approval before consequential actions, idempotent operations, a step and spend budget, full logs, an evaluation set with a pass rate, a fallback, and a cost per task below the value of the task. The Work Trend Index found 66 percent of leaders unwilling to hire without AI skills; agent teams want the person who adds the guardrails before the demo, not after the incident.

    What interviews test

    Expect to design an agent for a given workflow live - the loop, the tools, the permissions, the stopping rules - to debug a transcript of an agent going wrong, to discuss evaluation and cost, and to show one you built. Engineering roles add MCP, structured outputs and system design; product and operations roles add scoping and measurement. The strongest candidates can say what their agent is not allowed to do and how they know it is working.

    Proof they verify

    Hiring teams read the transcript and the evaluation: a repository with the loop, the tool definitions, the guardrails, the evaluation set and its results, plus logs from real runs, beats any demo video; an agent in use by someone else with adoption numbers beats a prototype. Pair it with a verified OneRoadmap AI Engineer certification and a short account of your own AI workflow.

    What gets you hired

    A working agent in use by someone, its evaluation and cost documented, a repository a reviewer can read, and a clear account of what it is not allowed to do. The sub-nodes give the ninety-day plan, the projects that get read and the proof that counts.

    A 90-day plan

    Month one: the agent loop built by hand against a model API with two tools and a stopping rule, plus the prompt and structured-output basics. Month two: guardrails - permissions, approvals, budgets, idempotency, logs - and an evaluation set of real tasks with a pass rate, on an agent that does one job for one real user. Month three: an MCP server for a real system, cost per task measured, a write-up of failures and fixes, the AI Engineer certification, and twenty targeted applications a week. Gartner expects a third of enterprise applications to include agentic AI by 2028, up from under one percent in 2024; a bounded, measured agent is what that demand wants.

    Projects that get read

    Projects that get read are bounded agents in use: a support triage agent that drafts and escalates with approval and logs; a research agent that produces a weekly brief with sources and a cost note; a bookkeeping assistant that categorises and proposes entries for an accountant to approve; an MCP server that exposes a real internal system safely. Each has a repository, an evaluation and a readme on what it may not do. Gartner's cancellation forecast is the market asking for exactly this discipline.

    Certification and proof

    Proof is verifiable: an agent someone uses, with its evaluation and logs public where they can be, a GitHub profile with the loop and the guardrails, a verified certification and a clear account of your workflow. With agentic skills up more than 280 percent in US postings in a year, the candidates who stand out are the ones whose agent is bounded, measured and running.

  2. 2. What agents are - and when not to build one.

    Agent Fundamentals

    What agents are - and when not to build one.

    Time: about 1 week. Topics: What makes an agent · Workflows vs agents · Agent use cases.

    What makes an agent

    Optional on the Experienced track

    Model + tools + loop: perceiving, deciding and acting toward a goal - versus a single prompt-response.

    Workflows vs agents

    Optional on the Experienced track

    Deterministic pipelines beat autonomy for most tasks - the decision framework before any code.

    Agent use cases

    Optional on the Experienced track

    Research, coding, operations, support - where the loop pays for its complexity today.

  3. 3. The mechanics of action.

    Tools & Loops

    The mechanics of action.

    Time: about 2 weeks. Topics: Tool / function calling · The agent loop · ReAct & planning patterns · MCP.

    Tool / function calling

    Declaring tools, executing calls and returning results - the primitive underneath everything agentic.

    The agent loop

    Plan → act → observe with stop conditions, budgets and retries - unbounded loops are outages.

    ReAct & planning patterns

    Interleaved reasoning and action; upfront plans vs adaptive steps - matching pattern to task.

    MCP

    The Model Context Protocol: standard connectors between models, tools and data sources.

  4. 4. Long tasks, multiple actors.

    Memory & Orchestration

    Long tasks, multiple actors.

    Time: about 2 weeks. Topics: Memory & state · Multi-agent patterns · Human-in-the-loop.

    Memory & state

    Context windows, summarisation and external stores - coherence beyond the context limit.

    Multi-agent patterns

    Orchestrator-worker, judge panels and hand-offs - parallelism and checks, at coordination cost.

    Human-in-the-loop

    Approval gates for consequential actions - autonomy with a supervisor's signature.

  5. 5. Evaluation and guardrails.

    Ship Agents Safely

    Evaluation and guardrails.

    Time: about 1-2 weeks. Topics: Evaluating agents · Safety & sandboxing · Observability.

    Build: A bounded agent. Build an agent that researches a topic with 2–3 tools, cites sources, respects a budget and stops cleanly.

    Evaluating agents

    Optional on the Fresher track

    Task success rates on golden sets, trajectory review and regression suites - measured, not vibed.

    Safety & sandboxing

    Optional on the Fresher track

    Allow-listed tools, permission prompts and prompt-injection defenses for tool-using systems.

    Observability

    Optional on the Fresher track

    Tracing every step, tool call and token - debugging loops instead of shrugging at them.

Prove it as you go

Reading is half of it. The free, scored AI Engineer certification covers the ground this roadmap leads to, and the AI Engineer job simulation puts it to work on real tickets - evidence a recruiter can check. Mark topics done, learning or skipped as you go; your progress carries into the classic view.

Fresher or experienced?

Both tracks hold every topic. On the Fresher track the foundations are filled and the later stages are outlined - do them when a project needs them. On the Experienced track the foundations are outlined and the later stages are filled. The 2026 section at the top applies to both.

Prefer a list?

The classic view shows the same 4 stages as a sequential checklist with time estimates. Your marks are shared between the two views.

Frequently asked questions

What is an AI agent, exactly?

A loop: a model reads a goal and the current state, chooses a tool, the tool runs, the result goes back into the context, and the loop continues until the job is done or a limit is hit. Everything else - memory, planning, multi-agent systems - is built on that loop, and learning to build it by hand is how you debug what frameworks hide.

Do I need to be an AI engineer to build agents?

To build them into products, yes - the AI Engineer roadmap is the path. To deploy them inside workflows through automation platforms and agent builders, no; operations, support, finance and marketing teams do it now, and Microsoft's 2025 Work Trend Index found 46 percent of leaders already using agents to automate work. Both need the same judgement about scope and guardrails.

Why do so many agent projects fail?

Cost and trust. Gartner predicts more than 40 percent of agentic AI projects will be cancelled by the end of 2027 for unclear value and inadequate risk controls, even as it expects a third of enterprise applications to include agentic AI by 2028. Agents scoped to a job that pays off and bounded with approvals, budgets, logs and evaluations are the ones that survive.

Which framework should I learn?

None first. Build the loop with a model API and two tools, add the guardrails, write the evaluation; then pick a framework if a project needs it. Frameworks change quickly and hide what you need to debug - what was retrieved, what the model saw, why a tool was called.

What is MCP and do I need it?

The Model Context Protocol is the common way to expose tools, data and prompts to agents from any provider, with its own transports, authorization and security concerns. Yes, learn to build an MCP server for a real system and to describe tools precisely - an agent with a badly described tool will use it badly.

How long does it take to learn agents?

With programming and model-API basics in place, a few weeks to build a bounded agent by hand and a few months to make one reliable with guardrails and an evaluation. Nobody can promise a timeline. An agent someone uses, with its evaluation, logs and a statement of what it may not do, is the evidence.

Will AI agents replace the people who build them?

Agents replace repeatable work and create the job of scoping, bounding, measuring and running them. Lightcast counted roughly 90,000 US postings mentioning agentic AI skills in 2025, up more than 280 percent in a year. The engineer who ships guardrails before the demo is the one those postings are for.

How do I prove agent skills without job experience?

A bounded agent in use: a support triage agent that drafts and escalates with approval and logs, a research agent that produces a weekly brief with sources and a cost note, an MCP server that exposes a real system safely. A repository with the loop, the guardrails, the evaluation and real run logs, plus OneRoadmap's AI Engineer certification.