Everything AI agents,
for developers building them.
New to building agents? Don’t start with a framework. Start with a fitting. We’ll measure how you want to build, then tailor you to the right path.
$whoami
An agent decides its own next step.
This is the one distinction that matters before anything else. A workflow runs a fixed set of steps you wrote. An agent runs a loop: the model looks at the situation, picks a tool, acts, observes, and decides what to do next — until the goal is met.
The arrow loops back: the model is in control of the path. Powerful, but harder to predict and test.
Go deeper on the loop and its primitivesThree questions. One tailored path.
The right way to build an agent isn’t about what sounds most sophisticated. It’s about where ownership should sit: who builds, who runs, who decides. Answer three and we’ll stitch a recommendation.
?Who should write the agent code?
step 1 of 3
Four patterns. Pick where ownership sits.
Building an agent in 2026 isn’t one decision — it’s four different paths. They’re not a ranking, and they combine. Each card is a pattern with its own ownership signature.
Build it yourself
Code-first. You own the full stack.
Write the agent yourself: model calls, tool wiring, memory, the orchestration loop, deployment, monitoring. Frameworks give you blocks; the architecture is yours.
Build it with a coding agent
You decide the architecture. The agent writes the code.
Agentic engineering, not vibe coding. You make the architectural calls before the agent touches the keyboard; Claude Code or Cursor handles implementation. You review the decisions, not just whether it compiles.
Deploy an open-source agent
It already exists. Clone, configure, run.
The OSS agent space has matured. When something already does 80% of what you need, deploying it on your own infrastructure beats rebuilding it. The question is no longer "does it exist?" but "which one is worth deploying, and why?"
Use a managed service
Someone else runs the harness. You configure via API.
The path to watch most closely in 2026. Session persistence, the execution environment, retry logic — these are rarely your competitive advantage. A managed service wires them so you can spend engineering time on what makes your agent actually better.
Every agent is cut from four pieces.
Whichever path you take, these are the materials on the bench. Hover or tap each to see what it does.
The model
The reasoning engine. It reads context and decides what to do next. Everything else exists to feed it the right input and act on its output.
Don’t start from a blank canvas.
Two open-source starting points from the Agentailor workshop. Production-shaped, not toy demos. Clone one and start cutting.
A production-shaped LangGraph + Next.js starter: MCP tools, human-in-the-loop, streaming, persistence. Decoupled enough to swap the orchestration layer.
$git clone https://github.com/agentailor/fullstack-langgraph-nextjs-agentScaffold a production-ready MCP server in seconds — session management, OAuth, transports, debugging baked in. Stop shipping demos.
$npx create-mcp-serverAgent Briefings
Stay measured as the field moves.
A bi-weekly newsletter on building and scaling AI agents in production. Practitioner, opinionated, no hype. Read by developers across all four paths.
Subscribe on LinkedInBuilt and written by Ali Ibrahim. Find the workshop on GitHub.
$subscribe agent-briefings
→ fortnightly dispatches on agents in production.
→ frameworks, MCP, evals, managed services.
→ signal over hype.