...build an agent
Jump to navigation
Jump to search
| AI Solver map |
|---|
|
Your current page is in bold
0 · Start
1 · Predict & analyze
tables, numbers, events
Predict values
Predict categories Forecast time series Cluster / segment Find outliers Make recommendations Find key drivers 2 · Understand unstructured data
text, documents, images, audio
3 · Generate & converse
4 · Decide & act
5 · Build it right
applies to every path
|
YouTube ... Quora ...Google search ...Google News ...Bing News
- AI Solver ... Agents/Assistants ... Personal Companions ... ...answer questions from my data ... ...choose a model ... ...evaluate and trust AI
Use this page when you want AI to do things, not just answer: file a ticket, update a CRM, research a topic, fix code, process a claim. An agent is a model using tools in a loop toward a goal. Agents are powerful but slower, costlier and riskier than simpler designs, so add autonomy step by step.
Start here
- Can one well-designed model call do it (with retrieval or a single tool)? → Yes: stop here; that is not an agent, and that is good → ...answer questions from my data
- Are the steps known in advance? → build a workflow: your code controls the path, the model fills in each step
- Fixed sequence → prompt chaining (draft → check → translate)
- Different inputs need different handling → routing (easy questions to a small model, hard ones to a large model)
- Independent parts, or several opinions → parallelization (split the work, or vote)
- Subtasks depend on the input → orchestrator–workers (one model plans, others execute)
- Quality improves with critique against clear criteria → evaluator–optimizer loop
- Steps cannot be predicted, the task is valuable, and mistakes can be caught or undone? → autonomous agent: the model plans, calls tools, checks results and repeats. Add stop conditions, time and cost budgets, and human approval for risky actions
- Must it operate software that has no API? → computer-use or browser agents that click and type. They are slower and more brittle; use APIs or MCP connectors whenever they exist
- Is it software work? → coding agents (Claude Code, OpenAI Codex, GitHub Copilot coding agent, Cursor, Gemini CLI); give them tests to run and an AGENTS.md file with project instructions
- Is one agent failing because it has too many tools or domains? → multi-agent: a manager agent that delegates, or agents that hand off to each other; across vendors, connect them with the A2A protocol. Expect higher cost and harder debugging
Connect tools and knowledge
| Standard | What it does |
|---|---|
| MCP (Model Context Protocol) | A standard way to give any agent access to tools and data (files, databases, SaaS apps). Write a connector once; use it from many agents. |
| A2A (Agent2Agent) | Lets agents from different teams or vendors discover each other, delegate tasks and return results. |
| AGENTS.md | A plain Markdown file in a code repository that tells coding agents how to build, test and follow conventions. |
| Agent Skills | Folders of instructions and scripts an agent loads when a task needs them. |
MCP, AGENTS.md and the goose agent framework were placed under the Linux Foundation's Agentic AI Foundation in December 2025, and Google donated A2A to the same foundation in August 2026. Agentic AI Foundation
Build options (October 2026)
| If you want... | Consider |
|---|---|
| No-code or low-code agents | Microsoft Copilot Studio; Gemini Enterprise (agent designer); OpenAI Agent Builder; Salesforce Agentforce; n8n, Zapier, Make |
| A provider's SDK | OpenAI Agents SDK; Claude Agent SDK; Google Agent Development Kit (ADK); Microsoft Agent Framework (successor to AutoGen and Semantic Kernel) |
| A model-neutral framework | LangGraph, CrewAI, LlamaIndex, Pydantic AI, smolagents, Mastra |
| Managed hosting, identity and governance | Amazon Bedrock AgentCore; Google Gemini Enterprise Agent Platform (formerly Vertex AI); Microsoft Foundry Agent Service |
Choosing models for agents: build the first version with the most capable model to see what is possible, then swap smaller, faster models into the steps where your evals still pass. Small language models handle routing, extraction and tool-argument formatting well. See ...choose a model.
Real-world examples
- Customer service at scale — Klarna's OpenAI-powered assistant handled about 2.3 million conversations in its first month (most of its customer-service chats), work Klarna equated to about 700 full-time agents. In 2025 its CEO said the push had lowered quality and Klarna began hiring human agents again. Design the human handoff from the start. Klarna is hiring customer service agents again | Entrepreneur
- Multi-agent research — Anthropic's research feature uses a lead agent that plans and parallel subagents that search; it outperformed a single agent on its internal research evaluation but used many times more tokens than a normal chat. Multi-agent systems pay off only for valuable, parallelizable tasks. How we built our multi-agent research system | Anthropic
- Cautionary tale — in July 2025 a Replit coding agent ran destructive commands and deleted a production database during an explicit code freeze. Instructions in a prompt are not permissions: separate development from production and restrict what the agent's credentials can do. AI agent wipes production database | eWeek
Watch out
- Least privilege — give agents the narrowest credentials possible, sandbox code execution, and keep production out of reach unless approved.
- Prompt injection — web pages, emails and documents can carry instructions; never let untrusted content trigger sensitive actions without checks (see the OWASP Top 10 for LLM applications).
- Human approval for payments, deletions, external messages and other irreversible steps.
- Observability — log every model call and tool call; review traces of failures weekly.
- Budgets — cap steps, time and spend per task.
Measure success
Task success rate on a fixed test set, cost and time per completed task, tool-call accuracy, how often a human had to step in, and safety violations. Read the agent's step-by-step traces, not just final answers. More: ...evaluate and trust AI
Learn more
- Building effective agents | Anthropic
- A practical guide to building agents | OpenAI
- Choose a design pattern for your agentic AI system | Google Cloud
- AI agent orchestration patterns | Microsoft
- Model Context Protocol (MCP)
- OWASP Top 10 for LLM Applications
- Agents/Assistants ... Personal Companions
Where next? ◀ Generate & converse · ...answer questions from my data · ...choose a model · ...deploy AI · ▶ ...evaluate and trust AI