Guides
The operating layer, in plain terms
These guides cover the decisions around agent work: what to delegate, which tools to grant, how to read a failed run, and where a person stays accountable. For a specific product, browse the tools catalog or compare options by job in the category archives.
Accounts and the planned features are not open yet. See what is planned; every guide and the catalog are free to read.
Guide
10 Tasks AI Agents Handle Well (and 5 to Keep in Your Hands)
Where AI agents earn their keep in 2026 and where they still fall short. A decision grid for delegating checkable, low-cost work and keeping judgment, money, and public commitments in human hands.
2026-09-26
Guide
How to Write Instructions AI Agents Actually Follow
Techniques for writing agent instructions that hold under real runs. Includes five before-and-after rewrites, a reusable template, and a three-run test that proves an instruction works before real work depends on it.
2026-09-26
Guide
AI Agent vs. Chatbot vs. Script: What's the Difference?
Agents, chatbots, and scripts do different work. A plain comparison of what each one is, how it fails, what it costs, and a five-question framework for choosing the simplest tool that fits the task.
2026-09-26
Guide
The 7 Most Common AI Agent Failures (and How to Catch Them)
Seven failure modes that show up in production agent work, from confident wrong answers to runaway loops. For each one: what it looks like, why it happens, and the specific check that catches it.
2026-09-26
Guide
A Checklist Before Letting an AI Agent Touch Real Work
Twelve checks before an agent touches customers, money, or public communication. Covers least-privilege access, written instructions and limits, trial runs, action logs, and rollback, calibrated by risk tier.
2026-09-26
Guide
How Much Does It Cost to Run an AI Agent? The Full Picture
Where agent spend actually goes: tokens, tool calls, model choice, retries, and review time. Includes a worked cost worksheet and the controls that keep per-task cost under control.
2026-09-26
Guide
Managing Agent Memory and Context
How agent context actually works, what fills it fastest, and the patterns that keep long runs coherent: a written state file, checkpoint summaries, trimmed tool outputs, and memory separated by lifetime.
2026-09-29
Guide
Multi-Agent Workflows That Actually Work
When multiple agents help and when they add cost without adding value. Covers orchestrator, generator-critic, and pipeline patterns, plus the contracts, shared state, and budgets that make them work.
2026-09-29
Guide
Agent Tool Permissions Done Right
Every tool you grant an agent is risk you accept. How to size blast radius, tier permissions, gate irreversible actions, design tools for least privilege, and audit what agents actually did.
2026-09-29
Guide
How to Evaluate Agent Output Quality
Eyeballing a few outputs is not an evaluation strategy. How to write acceptance criteria before the run, build a small runnable eval set, sample production output, and track quality as a time series.
2026-09-29
Guide
Prompt Chaining for Complex Agent Tasks
One giant prompt produces failures that are hard to locate. How to decompose complex work into chained steps with defined handoffs, verify between steps, and catch errors while they are cheap to fix.
2026-09-29
Guide
Using Agents for Code Review
Code review is high-leverage agent work when it is set up correctly. What agents catch, what they miss or fake, how to calibrate them with real diffs, and where human review stays mandatory.
2026-09-29
Guide
When Not to Use AI Agents
Knowing when not to deploy an agent is a core operating skill. Where agents are overkill, where they are dangerous, the hidden costs of defaulting to them, and what to use instead.
2026-09-29
Guide
Agent Security Risks You Must Handle
Agents read untrusted content, run code, and call APIs with your credentials. The risks that materialize in production, from prompt injection to credential leakage, and the layered mitigations that contain them.
2026-09-29
Guide
Comparing AI Agent Frameworks
Framework marketing converges on the same claims. The dimensions that matter for choosing: control, observability, tooling, deployment, cost, and team fit, plus a 30-day evaluation plan and the build-versus-buy math.
2026-09-29
Guide
AI Agents for Customer Support
Support is the most deployed agent use case and the most visibly failed. How to ground agents in current knowledge, design escalation to humans, handle angry customers, and measure resolution quality instead of demo performance.
2026-09-29
Guide
Testing Strategies for AI Agents
Non-deterministic does not mean untestable. How to test agents at every level, from deterministic tool units to statistical evals, chaos tests, and production shadow mode.
2026-09-29
Guide
Building Agent Error Recovery
Agents fail. The engineering question is whether failure degrades gracefully or corrupts state. A failure taxonomy and the recovery patterns that contain it: classified retries, fallbacks, circuit breakers, and idempotent re-runs.
2026-09-29
Guide
Writing SOPs Your Agents Can Follow
Procedures written for humans fail agents at every assumed judgment call. How agent SOPs differ, the anatomy of a procedure an agent can execute, and how to test, version, and maintain them.
2026-09-29
Guide
Measuring AI Agent ROI
Most agent ROI claims count time saved and ignore the costs that decide the outcome. The full formula, the costs teams forget, how to measure time saved correctly, and when to expand, redesign, or retire a workflow.
2026-09-29
Guide
Handoffs Between Agents and Humans
Work gets dropped at the seams between agents and people. How to design handoffs in both directions: self-contained briefings, five-part status reports, escalation triggers, and trust levels calibrated on evidence.
2026-09-29
Guide
How to Choose an AI Agent for a Specific Task
A job-first framework for choosing an agent for one specific task. Write the job down, score candidates on five criteria, run a same-task test on your own inputs, and cost the winner before it touches real work.
2026-10-04
Guide
Agent Harness Engineering: Where Tool Use Succeeds or Fails
The model gets the credit and the blame. The harness decides what actually happens: which call runs, with what arguments, under whose permission, and what occurs when it fails. This guide defines the layer, its six jobs, and the build order, drawn from published production practice.
2026-10-08
Guide
Diagnosing Tool-Call Failures: Read the Run, Fix the Layer
An agent run failed and the transcript will not say why. A field method for separating six failure layers (arguments, tool, transport, context, permission, sequencing) from per-call records, with the fix that belongs to each one.
2026-10-08
Guide
The Production Harness Checklist: 24 Checks Before an Agent Touches Real Work
Twenty-four pass-or-fail checks across the call, the gate, the run, and go-live. Each check requires evidence, not testimony, and traces to Anthropic engineering posts, the MCP specification, or LangGraph and OpenAI documentation. Widen scope only where every check holds.
2026-10-08
Guide
How AI Agents Actually Use Tools: Function Calling and MCP
A builder's guide to agent tool calling: JSON tool schemas, the reason-call-observe loop, and the MCP standard with transports and security.
2026-10-10
Guide
Designing Agent Actions That Can Be Undone: A Reversibility Plan Before You Automate
Classify every agent action by how hard it is to undo, pre-write the compensating action, add dry-run previews and idempotency, and scale approval gates to the tier before the agent runs on real work.
2026-10-13
Guide
Agent Observability: Log, Trace, and Read a Run After It Fails
An agent run is a tree of model calls and tool calls, and most teams keep none of it. What to record per run, how traces and spans fit together, what never belongs in a log, and when a platform earns its place over plain files.
2026-10-14
Guide
LangChain vs LlamaIndex vs Haystack for Building AI Agents
LangChain, LlamaIndex, and Haystack compared from their AgentsUse profiles, verified 2026-10-15: recorded strengths, limits, licenses, packages, MCP support, tool pairings, and which job each framework fits.
2026-10-17
Guide
What Is MCP (Model Context Protocol)? How Agents Connect to Tools
MCP explained from its own specification: hosts, clients, and servers; tools, resources, and prompts; stdio and Streamable HTTP transports; a first-server walkthrough; and how to judge a server before you install it.
2026-10-17
Guide
How to Choose a Browser Automation Tool for an AI Agent: Playwright, Puppeteer, Browser Use, or Stagehand
A concrete choosing guide for the browser layer of an agent, built only from the four published AgentsUse profiles in the Browser Automation category. Who decides the next click, which engines and languages each tool covers, how each connects to MCP, what it costs to run, and a three-task trial that settles the choice on your own pages.
2026-10-18
Guide
MCP Spec 2026-07-28: What Changed From 2025-11-25
A plain-English summary of the MCP 2026-07-28 changelog: stateless requests, server discovery, caching hints, deprecations, and what server users should check.
2026-10-29
Guide
MCP Security Best Practices in Plain English
A plain-English reading of the official MCP security page: local servers, token passthrough, scopes and SSRF, plus what to check before installing a server.
2026-10-29
Guide
Reading an MCP Tool Definition Before You Trust It
What the MCP specification says a server's tool definition contains, which parts are untrusted, how errors work, and a short checklist for reviewing a tool.
2026-10-29
Guide
MCP vs Direct API Integration: When to Use Each to Connect an Agent to a Tool
Direct function calling or an MCP server: how each one hands tools to an agent, what each costs in control and upkeep, and the four questions that pick between them.
2026-11-07
Guide
How to Build Your First MCP Server
Build a working MCP server from scratch in Python or TypeScript: one tool, one resource, tested in the Inspector, then connected to Claude for Desktop.
2026-11-09
Guide
How to Host an MCP Server (and When Not To)
Take a working stdio MCP server and put it on the network with Streamable HTTP: the transport switch, the auth decision, and the checks to run before you share the URL. Built from the 2026-07-28 specification and the official SDKs.
2026-11-11