WEBVTT 00:00:00.250 --> 00:00:04.016 An OmniNode is one node design, built for many kinds of jobs. 00:00:04.046 --> 00:00:07.605 It lives in Fabric Intelligence, inside Teleion Fabric. 00:00:07.959 --> 00:00:12.598 It picks up a Durable Task, and uses Durable Task Objects to do the work. 00:00:12.858 --> 00:00:16.000 Then a separate worker audits and approves the result. 00:00:16.030 --> 00:00:17.892 This design has six parts. 00:00:17.922 --> 00:00:18.554 Tools. 00:00:18.584 --> 00:00:19.204 Skills. 00:00:19.234 --> 00:00:20.203 A model. 00:00:20.233 --> 00:00:22.201 Hooks, such as a spending limit. 00:00:22.334 --> 00:00:23.291 Working files. 00:00:23.321 --> 00:00:24.488 And settings. 00:00:24.935 --> 00:00:28.076 For a new job, you change the parts, not the design. 00:00:28.170 --> 00:00:30.381 You can place it anywhere in a workflow. 00:00:30.411 --> 00:00:32.738 At the entry, to sort requests. 00:00:32.930 --> 00:00:34.016 On a branch. 00:00:34.207 --> 00:00:36.477 At the end, as one small tool. 00:00:36.669 --> 00:00:38.810 Beside another node, to audit it. 00:00:38.840 --> 00:00:41.145 Even inside another OmniNode. 00:00:41.190 --> 00:00:44.342 OmniNodes also join together, in four ways. 00:00:44.534 --> 00:00:45.688 One after another. 00:00:45.718 --> 00:00:46.895 Side by side. 00:00:46.925 --> 00:00:49.020 A backup, if the first one fails. 00:00:49.050 --> 00:00:52.214 And a pair, where one works and the other audits. 00:00:52.475 --> 00:00:57.220 The group is one OmniNode too, so small parts can build large systems. 00:00:57.250 --> 00:00:59.985 In a run, it tries the simple ways first. 00:01:00.338 --> 00:01:02.272 Code, when the answer is known. 00:01:02.463 --> 00:01:06.161 Reuse, when the work is already done and the details haven't changed. 00:01:06.352 --> 00:01:08.599 A Meta model, when it takes judgment. 00:01:08.953 --> 00:01:12.840 Then a rule tests the answer against the item's own text. 00:01:12.870 --> 00:01:14.164 Now, a real one. 00:01:14.194 --> 00:01:18.565 This OmniNode does feed maintenance, and a feed is simply a type of Task. 00:01:18.919 --> 00:01:24.347 Every four hours, it searches public GitHub for recently updated skill packs and tools. 00:01:24.608 --> 00:01:29.258 Muse Spark or Muse Wonder labels the new or changed ones. 00:01:29.704 --> 00:01:34.204 Once approved, each becomes a Task Object that any agent can find. 00:01:34.270 --> 00:01:36.993 Here are its first six recorded runs. 00:01:37.126 --> 00:01:41.556 Each looks at twenty-four items, and seven always stop at the license check. 00:01:41.816 --> 00:01:44.377 In run one, nothing can be reused. 00:01:44.452 --> 00:01:47.198 Eight items go to a model, and nine wait. 00:01:47.389 --> 00:01:50.147 In run two, four answers are reused. 00:01:50.280 --> 00:01:52.144 In run three, twelve. 00:01:52.404 --> 00:01:55.857 In run four, fifteen are reused, two wait for a retry, 00:01:55.887 --> 00:01:59.156 and no model is called at all: zero tokens. 00:01:59.510 --> 00:02:04.961 After that, a model sees only what needs it: two retries, then one project new to the feed. 00:02:05.408 --> 00:02:08.884 Tokens go from about nine thousand to under two thousand. 00:02:09.145 --> 00:02:14.644 Most listings stay the same between runs, so this is a real trace, not a savings claim. 00:02:14.750 --> 00:02:19.540 A separate worker audits and approves, with its own key and the other Muse model. 00:02:19.800 --> 00:02:22.848 It approved fifteen items, and rejected three. 00:02:23.109 --> 00:02:26.424 One model call failed, and it stays on the record. 00:02:26.490 --> 00:02:29.271 Then it learns from recorded runs and traces. 00:02:29.404 --> 00:02:32.290 After each run, a tuner changes the spec. 00:02:32.481 --> 00:02:37.259 It moved the model mix from fifty-fifty to about sixty-four, thirty-six. 00:02:37.334 --> 00:02:41.845 It raised the batch size from eight to ten, after a run passed its gate. 00:02:42.106 --> 00:02:46.840 With twenty outcomes, this is an early reading, not a model comparison. 00:02:46.870 --> 00:02:50.336 One design, built for many jobs, with every run on the record. 00:02:50.690 --> 00:02:54.225 Feed maintenance is our first OmniNode on its own timer. 00:02:54.486 --> 00:02:57.708 Connect your agents at teleionlabs.com.