Product placement
seanchen.io (acts: website where creator shares content, linked in bio, and hosts community; removing it would change what happens in the video)nGitHub (acts: platform where the agent's code is hosted and can be cloned; removing it would change what happens in the video)
On-screen text
Loop VS Graph Engineeringnseanchen.ion@SeanAIStories @ShenSeanChen @肖恩君SeannAI Agent Engineering LaddernPromptnRole PlayingnContextnData + PromptnSKILLSnProcedure Followingnso: write the skill first. harden it into a graph only once it stops changing.nLoopnbetween the callsnGoal-Driven IterationnGraphnshape of the callsnProcedure Pre-DeterminednTHE LOOP - it comes backnand look again... and againnyounLLMntoolnneeded?nyesnchoose & run antoolnnonreplynyou cannot draw where this ends, because it does not know either.nTHE GRAPH - it fans outnSTARTngithubnwebncalendarnmemorynsynthesizenrouterndraft_digestnanythingnto do?nnonENDnyesndraft itnevery edge is on the page before anything runs.nISN'T THIS JUST DETERMINISTIC WORKFLOWS FROM 2023?nMostly yes. Airflow has run conditional parallel DAGs for a decade.nWhat is new is narrow:n1. the nodes are non-deterministic — a node is a model calln2. a model can pick the edge — routing is data-dependentn3. so you need guards a DAG scheduler never didnOld shape. New contents.nLOOPn"work this PR"nstep 3 depends onnwhat step 2 found,nyou cannot draw it.nGRAPHn"brief me daily"nyou know the fournsources before younwake up.nTHE LOOP - the model decides what to call, one step at a timendecidengh gh calendar synthesizen7357msn3 sources. It never searched the web or memory — it decided it had enough.nTHE GRAPH - the shape is known, so the scans go togetherngithubn1506msnwebn1987msncalendarn2msnmemoryn1msnsynthesizen11279msnTHE BARRIER - a wave costs its SLOWEST branch,nso calendar and memory idle here for 1.9snISN'T THIS JUST DETERMINISTIC WORKFLOWS FROM 2023?nMostly yes. Airflow has run conditional parallel DAGs for a decade.nWhat is new is narrow:n1. the nodes are non-deterministic — a node is a model calln2. a model can pick the edge — routing is data-dependentn3. so you need guards a DAG scheduler never didnOld shape. New contents.nLOOPn"work this PR"nstep 3 depends onnwhat step 2 found,nyou cannot draw it.nGRAPHn"brief me daily"nyou know the fournsources before younwake up.nSEANCHEN.IOn@ShenSeanChenn@SeanAIStoriesn@ShenSeanChenn@肖恩君Sean