Guides

How to tell whether your task needs one well written prompt or something that can call tools, react to results, repeat on a schedule and remember.

An agent is the right build when you cannot list the steps in advance, because the next action depends on what the last one returned.

If you can draw the whole process as a flowchart with every branch labelled, you want a workflow, not an agent. Here is how to tell.

A pipeline moves a body of records through fixed transformations on a schedule, and its correctness question is completeness, not routing.

Multi agent systems cost far more tokens than a single agent, so use one only for real specialisation, real parallelism, or a checker that catches misses.

An MCP tool does one job when asked. If your description contains a decision the model must make, or more than one job, you are describing an agent.

Workflow, agent and pipeline are not interchangeable. The difference is what decides the next step: a branch you wrote, the model, or the schedule.

Anthropic calls context engineering the natural progression of prompt engineering. Here is what that actually changes in your work.

Karpathy named vibe coding in 2025. The tooling reaction since is uniform: write the intended behaviour down before the model writes code.

A Skill changes what the model knows how to do, an MCP server changes what it can reach, and a subagent changes whose context does the work.

The six ways AI agents break in production, what each one looks like in a real incident, and the specific spec line that would have caught it.

Why an agent that behaved for ten turns stops behaving at turn forty, and what the attention budget and context rot research actually measured.

How to pick a complexity tier, an autonomy level and a memory type when the signals are ambiguous, and why rounding down beats rounding up.

A real tool fallback names the retry budget, what happens when it runs out, and which of four outcomes applies. Handle the error is not one.

A plain-English guide to the Model Context Protocol at revision 2026-07-28: what a server offers, what clients do, and what older guides get wrong.

A section by section walkthrough that turns one sentence about a support triage agent into a working agent spec, and what each part prevents.

Two agents write the same field and the last writer wins silently. The fix is a declared merge strategy per field, not more coordination.

In common usage the terms are interchangeable, but a workflow and a pipeline fail in different ways, which is why they need different specs.

Why a bare prohibition fails under a competing instruction, the four guardrail classes an agent spec needs, and when to require approval.

A prompt is consumed and gone. A spec persists, so intent can be reviewed, diffed and argued with before your coding agent writes a line of code.