Frequently asked questions

A chatbot answers; an agent acts. Agents use the model to decide their next step — calling tools, reading data, checking their own work — in a loop, until the goal is done. The chat interface may look identical; the loop underneath is the difference. Build the loop in 30 minutes →

No — the core agent loop is about 40 lines of code, and for one agent with a handful of tools, rolling your own is less complexity, not more. Frameworks earn their keep at durable long-running state and parallel multi-agent graphs. Build the bare loop once first, then decide. The build-vs-buy verdict →

No. MCP is an open standard with clients across the ecosystem — Claude apps and Claude Code popularized it, but many IDEs, agent frameworks, and model providers speak it. A server you write once works in any compliant client. Build your first server →

The stable answer is a method, not a model name: route by task tier — a frontier model (Claude, GPT, Gemini class) for planning, cheaper or open-weights models for routine steps — and let golden-task evals on your own workload pick the names. Leaderboards measure chat, not multi-step tool reliability. The six criteria that matter →

Yes — open-weights models served by llama.cpp or vLLM expose an OpenAI-compatible endpoint, and the rest of the agent stack (harness, MCP tools, evals) is just your code. Expect local 7–14B models to be weaker at long-horizon planning than frontier hosted models; measure on your tasks. Local serving guide →

They can be — if safety comes from architecture rather than prompts. The controls that work: capability allowlists at the tool layer, approval gates for irreversible actions, trajectory logging, and evals run on every change. "We told the model to be careful" is not a control. The working threat model →

Agent cost = usage × verbosity × retries × context length — four multipliers one commit can change. Meter per trajectory (not per call), route task tiers to appropriately-priced models, and cap retries in the harness. Instrument before you scale; dashboards of regrets are expensive. FinOps for agents →

Test outcomes and trajectories, not phrasing: golden tasks with checks like "the answer contains the right fact", "the right tool was called", "the forbidden tool was not". Run tasks multiple times, gate CI on pass rates, and turn every production incident into a new golden task. Build the eval harness →

It is malicious instructions hidden in content your agent reads — a web page, an email, a document. If your agent has tools and touches untrusted content, yes: assume injection sometimes succeeds and design so a successful one cannot reach anything worth the attack. OWASP now maintains a peer-reviewed top-10 for agentic applications. Defenses that actually work →

Use the Start Here page — pick a goal and get a numbered path. If you just want the fastest meaningful win: build the 40-line agent loop, then your first MCP server. Both run on a laptop with no API key. Start Here →