Agentic AI Course: How to Build AI Agents, Beginner to Advanced (in PHP)

Before you start

• PHP 8.1 or newer with Composer and the cURL extension — the free local setup is covered in lesson 2
• You can call a web API and read JSON; the AI for PHP Developers course is the ideal warm-up but is not required
• An Anthropic API key on a pay-as-you-go plan — a few dollars covers every exercise in the course

Curriculum

Lesson 1 is free to read, no account needed. The rest unlock the moment you buy.

1. 1 How to build an AI agent — what one is, and when not to build one Preview free What an AI agent actually is (a model, tools and a loop), how it differs from a plain API call or a workflow, the four ways to build one, and the four questions to ask before you do. 4 min
2. 2 Set up on localhost — the official PHP SDK and your first call Install the Anthropic PHP SDK with Composer, keep your API key out of code, make your first call, read the response safely, and understand what a call costs. 3 min
3. 3 Your first tool — how the model asks your PHP to do something Define a tool with a name, a prescriptive description and a JSON schema; see exactly what the model sends back; run the tool and return the result; and write the simplest agent loop by hand. 3 min
4. 4 The loop, done properly — stop reasons, errors, limits and parallel calls Handle every way a turn can end, return tool errors the model can recover from, cap the number of turns, run parallel tool calls correctly, and log every turn. 3 min
5. 5 The Tool Runner — the same agent in a fraction of the code Let the official PHP SDK run the loop for you with BetaRunnableTool and toolRunner(); learn what it does and doesn't do, where to put approval gates, and when the manual loop is still the right choice. 3 min
6. 6 Designing tools that work — descriptions, schemas and what to expose The rules for tool design: prescriptive descriptions, enums and required fields, strict schemas, a focused tool set, and the bash-versus-dedicated-tool decision that decides what your harness can control. 4 min
7. 7 Reliable output — getting exactly the JSON your code needs Use structured outputs to get the agent's final answer as validated JSON, in a PHP class or a raw schema; know the schema limits; and handle the model rules that replaced forced tool use and prefilling. 3 min
8. 8 Project — the support-desk agent, end to end Put the Beginner tier together into one working agent that looks up orders, checks stock, drafts a reply for a human to approve, and returns a typed result. 3 min
9. 9 Rules and guardrails — allow, ask, deny Give every tool a permission policy, put a human in front of irreversible actions, cap turns and tokens, and build a kill switch — the rules that make an agent safe to run. 3 min
10. 10 Security for agents — untrusted input, paths, shells and secrets Why tool results are untrusted input, how indirect prompt injection works, confining file paths, never giving an agent a free shell, and keeping secrets out of prompts, memory and logs. 3 min
11. 11 Cost control — effort, the right model per job, and spend caps Understand what an agent run costs, tune effort, use a cheaper model where it holds quality, count tokens before you send, log usage per run, and enforce daily caps. 3 min
12. 12 Prompt caching for agents — the biggest cost lever How prompt caching works (a prefix match), where to put cache breakpoints in an agent, how to verify hits, and the silent mistakes that turn caching off without any error. 3 min
13. 13 Memory — agents that remember across sessions The difference between conversation state and long-term memory, implementing the memory tool in PHP with strict path safety, and what must never be stored. 3 min
14. 14 Long-running agents — context editing, compaction and memory What happens when an agent's conversation grows too large, and the three tools for it — clearing stale tool results, server-side compaction, and memory — with when to use each. 3 min
15. 15 Server tools — web search, web fetch and code execution Use tools that run on Anthropic's servers: web search and fetch with domain limits and citations, the code-execution sandbox, handling pause_turn, and why server-tool errors don't raise exceptions. 3 min
16. 16 Project — a research-and-report agent with guardrails Combine the Pro tier into one agent that researches a topic on trusted sites, stays inside a budget, caches its prompt, logs cost, and produces a cited report that a human approves before it's published. 3 min
17. 17 MCP — connecting agents to other people's tools What the Model Context Protocol is, how to connect a remote MCP server with the MCP connector (both required halves), allow-listing its tools, and when MCP beats writing your own tool. 3 min
18. 18 Many tools — tool search, skills and changing the tool set What to do when an agent needs dozens or hundreds of tools: load them on demand with tool search, package instructions as skills, and change the tool set mid-conversation without breaking the cache. 3 min
19. 19 Multi-agent — an orchestrator and workers When one agent should hand work to others, how to build an orchestrator with cheaper worker calls in PHP, why workers return results rather than transcripts, and the caching and cost trade-offs. 3 min
20. 20 Managed Agents — letting Anthropic run the loop and the sandbox What Managed Agents are, the agent-once/session-every-run model, streaming session events in PHP, permission policies, vaults for credentials, schedules and budgets — and when to choose it over self-hosting. 3 min
21. 21 Streaming, UX and evals — agents people trust Stream agent progress to the browser, show which tools are running, handle large streamed tool inputs safely, and build a tiny eval set so you know whether a change made the agent better. 3 min
22. 22 Hosting — running an agent on ordinary shared PHP hosting Why agents clash with shared hosting's time limits, the job-queue pattern that fixes it (process a few jobs per run, reconnect the database, requeue stuck jobs), cron, secrets on the server, and when you need a VPS. 3 min
23. 23 The production checklist — shipping an agent you can sleep through Retries and timeouts, the right error-handling chain, refusals, versioning prompts and tools, monitoring usage and quality, rate limits, and a rollback plan — as a checklist. 3 min
24. 24 Capstone — design and ship your own agent, and what's next Three capstone briefs to build end to end with everything in the course, the honest limits of agents, how to get your certificate of completion, and where to go next. 3 min

How it works

1. 1 · Buy once

   One payment in your own currency, no subscription. Your account is created from the email you pay with.

2. 2 · Log in and read

   Every lesson opens in your account on any device. Mark each one done and pick up where you left off.

3. 3 · Keep the Word file

   The whole course as a .docx — annotate it, print it, read it offline. Re-download it any time from your orders.

Where this course fits

1. AI for PHP Developers
2. →
3. Agentic AI Course · you are here
4. →
5. Advanced PHP Architecture

Questions about this course