AI Agents Cheat Sheet

By the Syncfusion DevTutorials team · Last updated October 19, 2026

Every key idea from the AI Agents course on one page: the loop, stop rules, permissions, planning strategies, memory, guards, dev-workflow agent configs, metric formulas and an error index.

Everything from this course on one page. Use it as a quick reference while building your own agents. Each section points back to the lesson where the topic is explained in detail.

The Agent Loop

The loop from Lessons 1, 3, and 9: look → think → act → check. Repeat until a stop condition says the task is finished.

The loop template

async function runAgent({ goal, brain, tools, maxSteps }) {
  const steps = [];
  for (let n = 1; n <= maxSteps; n++) {
    const action = await brain(goal, steps);                 // think
    if (action.final) return { stop: "final", answer: action.final, steps };
    const tool = tools[action.tool];
    const observation = tool ? await tool(action.args) : "ERROR: unknown tool " + action.tool; // act
    steps.push({ n, action, observation });                  // check: the next think reads this
  }
  return { stop: "max-steps", steps };
}

const tools = { run_tests: async () => "3 passed, 0 failed" };
const brain = async (goal, steps) => steps.length ? { final: "All tests pass." } : { tool: "run_tests" };
runAgent({ goal: "Make the tests pass", brain, tools, maxSteps: 5 }).then((r) => console.log(r.stop + ": " + r.answer + " (" + r.steps.length + " tool step)"));

Expected output:

final: All tests pass. (1 tool step)

Agent or Workflow?

Stop Rules and Budgets

Boundaries and Permissions Checklist

• write_file may only change files in the agent's own folder. (Lessons 1, 5)
• Test files are forbidden for agents that fix code. (Lessons 5, 19)
• Every tool call is checked against its schema before it runs. (Lesson 11)
• Each agent has only the tools its role needs. (Lessons 11, 24)
• Read-only mode for agents that only inspect. (Lessons 11, 24)
• A file must be read before it is written. (Lesson 12)
• Risky actions require a person's approval, one approval per action. (Lesson 24)
• Tool results are data, never instructions. (Lesson 24)

Planning Strategies

Memory Techniques

Guards

Dev-Workflow Agent Configs

Metric Formulas

The four core metrics

const metrics = {
  toolSuccess: (r) => r.toolOk / r.toolCalls,
  decisionAccuracy: (r) => r.passed / r.tasks,
  planCompletion: (r) => r.verified / r.tasks,
  recoveryRate: (r) => r.faultyPassed / r.faulty,
};
const baseline = { toolOk: 26, toolCalls: 33, passed: 6, verified: 5, tasks: 10, faultyPassed: 2, faulty: 6 };

for (const [name, f] of Object.entries(metrics)) console.log(name.padEnd(17) + Math.round(f(baseline) * 100) + "%");

Expected output:

toolSuccess      79%
decisionAccuracy 60%
planCompletion   50%
recoveryRate     33%

These are the baseline agent's numbers from Lesson 25's mini-benchmark.

From Lab to Production

The full table with common traps is in Lesson 25.

Error Index

Thanks for completing AI Agents. A good next step is the Agent Skills course, which packages what an agent knows into reusable skills.