Jake Lawrence · Information Visualization / AI Systems / HCI · AI Systems theme
One synthetic four-agent run goes wrong in a planted way: a source check is skipped, an uncited claim slips through, and the critic accepts it with every span green. Eight visualization techniques from the literature each show a different slice of that failure, live, and two cannot see it at all.
A long-form post that walks through eight techniques for visualizing LLM multi-agent pipelines (span waterfalls, execution graphs, hierarchical temporal summaries, layered semantic summaries, branching message histories, coordination plans, cross-run comparison, and failure-mode overlays), each with a working demo inline. Every demo renders the same synthetic run, shaped like OpenTelemetry GenAI agent spans, so the reader sees one set of facts eight ways and the comparison becomes the argument. The run is planted: an orchestrator plans a source check that no agent owns, a researcher returns a claim with no citation, a writer loops once under a critic that rejects for an unrelated reason and then accepts, and every span reports success. Which views expose the skipped step, the uncited claim and the incorrect verification, and which flatten them away, is the reading. Closes with a loop-native thesis demo (the run drawn as laps on a closed track) and the gaps a literature review of sixteen sources found most promising. No runtime model calls; everything is pre-computed in a committed fixture and re-verified in CI. A companion review page carries the full source ledger.
SAGEN argues an agent has no persistent model of where it is or what it has done. Eight Ways to See an Agent Think is the observer's side of the same gap: the traces exist, every span is green, and the view still has to be built that shows a step was skipped.
Classification as Infrastructure says every planning layer is a classification system whose consequences stay invisible. The failure-mode overlay in Eight Ways makes that classification visible on the trace itself, and finds the largest planted failure sitting in the plan, not in any tool call.
AlphaFold publishes, residue by residue, exactly where not to trust it. An agent trace does the opposite: its structural status is green everywhere, including on the step that let an unverified claim through. Eight Ways is about the views that could draw the missing confidence back on.
This is field data, not commentary: dated, sourced, and versioned. Subscribe and the next release, correction or new thread reaches you when it lands. It is the one list for the whole site, so you will get the rest of the work too. No schedule, easy out.
Back to the release table · See it on the network · Machine-readable catalog
I design and ship AI tools, full-stack apps, and data pipelines — end to end, to production. Tell me the problem in a sentence; I'll give you an honest read on fit within a day.
Work with me →