Agent Engineering

You build an agent when a task repeats, needs tools, and does not fit inside an interactive session: a PR reviewer, a nightly dependency updater, a support triage bot, a data-migration assistant. Pick the lowest layer that does the job.

The Claude Agent SDK exposes the same agent loop, tools, and context management that power Claude Code, in Python (claude-agent-sdk) and TypeScript (@anthropic-ai/claude-agent-sdk). query() runs a task and streams messages; ClaudeSDKClient keeps a session across turns and supports interrupts.

import { query } from "@anthropic-ai/claude-agent-sdk";

for await (const message of query({
  prompt: "Where do we validate session tokens? Cite files.",
  options: { allowedTools: ["Read", "Grep", "Glob"] },
})) {
  if (message.type === "result") console.log(message.result);
}

• allowedTools is your first guardrail — start read-only and add Edit/Bash only when needed
• Hooks, custom tools, MCP servers, subagents, and skills are all configurable in options
• Auth via ANTHROPIC_API_KEY, or Bedrock/Vertex/Foundry through environment settings
• For other languages, run the CLI with -p and --output-format json

The OpenAI Agents SDK (Python 'openai-agents', TypeScript '@openai/agents') is a lightweight framework for tool-calling agents. Its core ideas: Agents (instructions + tools), a Runner that executes the loop, handoffs to specialist agents, guardrails that validate input and output, sessions for memory, and built-in tracing.

from agents import Agent, Runner, function_tool

@function_tool
def lookup_order(order_id: str) -> str:
    return orders.get(order_id)

billing = Agent(name="Billing", instructions="Resolve billing issues.",
                tools=[lookup_order])
triage  = Agent(name="Triage",  instructions="Route to the right specialist.",
                handoffs=[billing])

result = Runner.run_sync(triage, "I was charged twice for order 812")
print(result.final_output)

• Handoffs keep each agent's instructions short and focused
• Guardrails run checks in parallel and can stop a run early (tripwires)
• Traces show every model call, tool call, and handoff — use them to debug and to build eval sets

The Cursor SDK (TypeScript '@cursor/sdk', Python 'cursor-sdk', public beta) runs Cursor agents from your own code, locally against a directory or in the cloud against a cloned repo that can end in a PR. Agent.prompt() is a one-shot; Agent.create() + agent.send() supports streaming and follow-ups; Agent.resume() continues an agent later.

import { Agent } from "@cursor/sdk";

const result = await Agent.prompt("Fix the failing lint errors in src/", {
  apiKey: process.env.CURSOR_API_KEY!,
  model: { id: "composer-2.5" },
  local: { cwd: process.cwd() },
});
process.exit(result.status === "finished" ? 0 : 2);

• Always set local or cloud explicitly — local is the silent default
• A thrown CursorAgentError means the run never started; result.status === 'error' means it ran and failed
• Dispose agents (await using / with) so long-running services do not leak processes
• Cloud runs can open PRs automatically; a REST API covers other languages

Some tasks live behind a UI: reproducing a front-end bug, verifying a checkout flow, filling a legacy admin form. Browser and computer-use tools let agents see and act on screens.

MCP connects an agent to tools and data. A2A (Agent2Agent) connects agents to other agents across frameworks and vendors. Contributed by Google and now a Linux Foundation project, A2A reached v1.0 in 2026: agents publish an Agent Card at /.well-known/agent-card.json describing skills, endpoint, and auth, then exchange Tasks and Messages over JSON-RPC with streaming and push notifications.

• Operate agents like services: timeouts, turn caps, budgets, retries only when safe
• Trace every run (LangSmith, OpenAI tracing, harness logs) and keep IDs in your logs
• Version prompts, tools, and models together; eval before rollout
• Least-privilege credentials per agent; humans approve irreversible actions

A team adds an AI reviewer to every PR.

bot: full write access, no turn limit, prompt "review this PR"
PR #481: 23 comments — naming nits, 'consider adding docs',
         one real bug buried at #17
team: mutes the bot within a week

script: Claude Agent SDK, allowedTools [Read, Grep, Glob], maxTurns 12
skill:  team checklist (security, error paths, tests, migrations)
rule:   max 5 findings, file:line + severity + fix, else "LGTM"
CI:     one sticky comment; skipped for fork PRs
metric: 2 weeks → 61% of findings acted on

• Least privilege and a turn cap bounded risk and cost
• A checklist skill replaced a vague 'review this' prompt
• A cap on findings forced prioritization; a metric proved value