The Claude agent loop explained — Transcript

Full transcript

1. 0:02You've made API calls, but a single call
2. 0:04only returns [music] one response.
3. 0:06If you want to replace workflows, well,
4. 0:08Claude needs to act, look at the result,
5. 0:11decide what's next, and keep going. And
6. 0:13this is what's commonly known as agentic
7. 0:15workflows.
8. 0:18An agent is an autonomous version of
9. 0:20Claude running both sides of the
10. 0:21messaging loop without human
11. 0:22involvement. Agents receive a task, pick
12. 0:25a tool, and execute code in a loop until
13. 0:28Claude deems the task done.
14. 0:31Here's the easiest way to implement an
15. 0:32agent loop.
16. 0:33First, send a message to Claude with
17. 0:35tools available.
18. 0:37Then, Claude responds with either a
19. 0:39final answer
20. 0:40or uses a tool that you define.
21. 0:43Then, your code executes that tool.
22. 0:46And then that result is sent back to
23. 0:48Claude.
24. 0:49And this repeats until stop reason is
25. 0:51end turn.
26. 0:55I want to see this loop run end to end
27. 0:56without dragging in a database or UI.
28. 0:58So, I'll wire up a fake Get weather, and
29. 1:01ask Claude what to wear in Austin today.
30. 1:04Claude has no way to know the weather on
31. 1:06its own, so it'll have to call the tool,
32. 1:08read the result, and then give you an
33. 1:09answer. So, here's the whole script.
34. 1:12First, the tools array tells Claude
35. 1:14what's available. Name,
36. 1:16description,
37. 1:18and a JSON schema for the inputs.
38. 1:21Run tool is just a hard-coded lookup. In
39. 1:24a real app, this would hit your
40. 1:25database, an API, whatever.
41. 1:28And then our loop here is the agent
42. 1:29loop.
43. 1:30Each iteration sends messages to Claude
44. 1:32and switches on the response stop
45. 1:34reason. And so, if the response is end
46. 1:37turn, then Claude is done. Print the
47. 1:39final text and just break. But, on tool
48. 1:42use, find the tool use block in the
49. 1:44response, run each one, and push the
50. 1:46assistant's response and our tool
51. 1:47results back into messages and loop
52. 1:49again, so Claude can answer.
53. 1:52So, let's run it.
54. 1:53You'll see two turns. Turn one is the
55. 1:56stop reason, which is tool use.
56. 1:58Claude requests get weather, which is
57. 2:00Austin in this case.
58. 2:02Our code returns the temp and the
59. 2:04conditions.
60. 2:06And then turn two, the stop reason is
61. 2:08end turn. And so Claude tells you to
62. 2:10wear something light and breathable. Two
63. 2:12API calls, one tool execution, and one
64. 2:14final answer. And that's the entire
65. 2:16loop. Everything that you make with the
66. 2:18Claude API is going to be similar to
67. 2:20this. And so in a real environment, this
68. 2:23same loop powers something like auto
69. 2:25review endpoint, a compliance agent that
70. 2:27reads a structural report, looks up the
71. 2:29relevant building codes via tool, and
72. 2:31writes risk finding back to the database
73. 2:33one by one as it works. The shape of the
74. 2:36loop is identical to what you just ran.
75. 2:38The differences are real tools instead
76. 2:40of a mock weather lookup, results stream
77. 2:42back to the UI as server-sent events,
78. 2:45and findings persisted to a risk finding
79. 2:48table.
80. 2:50An agent is Claude in a loop. Observe,
81. 2:53decide, act, repeat. You own the loop
82. 2:56and the tools. Claude owns the
83. 2:58reasoning. And when you don't want to
84. 3:00own the loop, managed agents run this
85. 3:02exact loop for you on Anthropic's
86. 3:04infrastructure.