This demo is a working example of what Etch Labs ships: expert systems with a modern agent interface. The chat feels like a frontier LLM agent — streaming thoughts, tool calls, structured handoffs — but the heavy lifting runs through deterministic Etch experts wired to real (mock) business APIs.
What you are looking at
Etch is the LLM-powered orchestrator you talk to. It interprets intent, chooses which expert to invoke, and narrates the result in plain language. Experts are separate programs built with the Etch SDK: fixed pipelines that call GLC microservices, aggregate data, and return auditable JSON. They stream the same UX primitives users expect from Claude or GPT — reasoning text, tool badges, progress metrics — without probabilistic generation inside the expert itself.
What experts are good at
- Grounded lookups — order status, invoice matching, inventory scans tied to live API records.
- Multi-step workflows — procure-to-pay and morning ops brief chain dozens of tool calls with predictable ordering.
- Structured handoffs — every expert returns JSON Etch can cite; nothing is invented.
- Speed at scale — typical runs finish in under two seconds with thousands of tokens of streamed reasoning.
- Auditability — expand any expert panel in chat and copy the full transcript; traces are reproducible.
What experts are not
Experts do not hallucinate facts they were not given — they cannot invent an order ID or invoice total. They can still make mistakes: bad source data, wrong parsing of an ambiguous email, or a business rule edge case the pipeline was not built for. The difference from an LLM is that every step is traceable and fixable in code rather than prompt-tuning roulette.
Try it
Use New chat and the quick pills — Order Tracking, Accounts Payable, Procurement, Operations, Sales — or paste your own request. Each pill triggers Etch, which delegates to the matching expert. See Two-layer architecture for how the pieces connect, and Example traces for full runs with token counts and tool timelines.