Moving to Mynd
Migration
Most teams arrive from one of three places: a raw LLM API with hand-rolled retrieval, an RPA stack, or a previous major version of Mynd. The shape of the work differs; the destination is the same — runs with traces.
From a raw LLM + DIY RAG
- Your retrieval pipeline collapses into context sources — ingest once, stop managing chunking and reranking
- Prompt templates become run prompts; output parsing becomes output_schema
- Map your retry/validation glue to max_steps and in-run schema retries
- Keep both systems live on the same eval suite for two weeks; cut over when the judge scores converge
From RPA
Each bot becomes an agent run with declared tools. The translation discipline: every action the bot performed silently becomes a scoped, audited tool call, and irreversible steps get approval gates. Expect to delete the screen-scraping half of the bot entirely — Y0 Vision reads the screens the bot was built to click through.
v1 → v2 API
| v1 | v2 | note |
|---|---|---|
| POST /v1/complete | POST /v1/runs | Completions are runs with max_steps: 1 |
| context_ids[] | context[] sources | The graph resolves items; explicit ids still work via items[] |
| steps (advisory) | max_steps (enforced) | Overruns now halt instead of billing through |
| X-Mynd-Key header | Authorization: Bearer | Old header accepted until 2026-12-31 |
Deadlines
v1 is in maintenance: security fixes only, sunset 2026-12-31. The CLI's `mynd migrate scan` greps a codebase for v1 calls and prints the v2 equivalent for each hit. Release notes track every deprecation with at least 12 months of runway.