What are managed agents on the Claude Platform? — Transcript

Full transcript

1. 0:00[music]
2. 0:02>> Claude managed agents is a suite of APIs
3. 0:04for building and deploying agents at
4. 0:06scale.
5. 0:11You define agents with specific tools,
6. 0:13personas, and capabilities. [music]
7. 0:15You configure sandbox environments with
8. 0:17the right packages and network controls.
9. 0:20You fire off sessions from [music] your
10. 0:22own application, and then Claude does
11. 0:24the work inside an isolated container
12. 0:26with full file system access, bash
13. 0:28execution, and web search. Under the
14. 0:31hood, this is an agent loop. Claude
15. 0:33reasons, calls a tool,
16. 0:35reads the result, and repeats until the
17. 0:37job is done. We've built this kind of
18. 0:39loop before ourselves, but managed
19. 0:41agents takes that same loop and hosts it
20. 0:43on Anthropic's infrastructure.
21. 0:46I have over here a Kanban board sitting
22. 0:48on top of managed agents.
23. 0:50I drag one over to the in progress, and
24. 0:52then that fires off a session
25. 0:54automatically. Now, the ticket says
26. 0:56optimize website performance. So, my
27. 0:58back end creates a session. It points to
28. 1:01an environment that I configured with
29. 1:03Lighthouse and Puppeteer pre-installed,
30. 1:06and mounts my GitHub repo into that
31. 1:08container. Claude has the code base, the
32. 1:10tools, and a rubric. Lighthouse score
33. 1:13above 90, no render-blocking resources,
34. 1:15all images lazy loaded. And then we can
35. 1:18see here that Claude runs the audit. It
36. 1:21starts compressing images, inlining CSS,
37. 1:23deferring scripts.
38. 1:25Every tool call streams back to the
39. 1:27board in real time through the event
40. 1:29stream.
41. 1:30So, the rubric kicks in. A separate
42. 1:32grader running at its own context window
43. 1:33evaluates the output against my
44. 1:35criteria. Claude reads that feedback,
45. 1:38goes back in, fixes what it misses, and
46. 1:40then resubmits.
47. 1:43Good. We're up to 96.
48. 1:45And note that I can drag a second ticket
49. 1:47over while the first is still running.
50. 1:49Two sessions, two containers, two
51. 1:52separate tasks running in parallel.
52. 1:55So, I have another agent here that's job
53. 1:57is to track prices and plan changes
54. 1:59across every SaaS tool that our company
55. 2:02pays for and have a report ready before
56. 2:05stand-up. Comment.
57. 2:07Claude searches the web for current
58. 2:09pricing pages, checks for plan tier
59. 2:11changes, flags new features that might
60. 2:12affect your contracts. It then runs a
61. 2:15cost analysis in Python inside of that
62. 2:17sandbox. And then it also uses an Excel
63. 2:19spreadsheet skill and writes an
64. 2:21executive summary. And when the report
65. 2:22is ready, Claude posts a link to Slack
66. 2:24and creates a review task in Asana, both
67. 2:27through MCP servers.
68. 2:29The agent also reads from and writes to
69. 2:31a memory store.
70. 2:33Before it starts, it checks what it
71. 2:35found last week. After it finishes, it
72. 2:37stores what changed. So, next Monday's
73. 2:39report says, "Cloud compute 15% lower
74. 2:42since last week." instead of listing the
75. 2:44same static pricing data.
76. 2:47I have an alert here that fired from my
77. 2:49monitoring stack. A custom tool my back
78. 2:52end receives the alert payload and sends
79. 2:54it into a new session as a tool result.
80. 2:56This session uses multi-agent
81. 2:58coordination. A coordinator agent
82. 3:00receives the alert and delegates to
83. 3:02three specialists, each running in their
84. 3:04own context window on the same shared
85. 3:06file system.
86. 3:08The specialists report back. The
87. 3:10coordinator synthesizes their findings
88. 3:12into a single incident summary. And
89. 3:14before it posts the update to Slack, the
90. 3:17permissions policy fires. And so, I see
91. 3:19the draft on screen, approve it, and the
92. 3:22message goes out. Memory ties all of
93. 3:24this together. The coordinator checks
94. 3:26past incidents in the memory store and
95. 3:28flags a pattern. This looks like the TDS
96. 3:31resolution issue from 2 weeks ago that
97. 3:33was caused by a misconfigured TTL. So,
98. 3:35that means that the next time a similar
99. 3:37alert fires, the agent starts with that
100. 3:39context instead of diagnosing from
101. 3:41scratch.
102. 3:46Managed agents gives developers [music]
103. 3:48the tools to deliver a fully managed
104. 3:49stateful agent experience with agents,
105. 3:52sessions, environments, tools, MCP,
106. 3:55memory, outcomes, [music] and
107. 3:56multi-agent coordination.
108. 3:59You define what done looks like. Claude
109. 4:01works until [music] it gets there.