MICROSOFT AGENT FRAMEWORK · 0.16.0-rc.1
Exercise guide
Choose completion, experimentation and correction reading from the actual distributed registry.
Objectives
- One Actionability Start/End pair
- Manipulations of complete programs
- Historical exercise correspondence
Learn from a falsifiable observation
Begin an exercise with a prediction: which state, error or row should exist? Then record command, profile, input and result. A text screenshot alone does not establish a business effect. Work from the complete distribution; use a copy for changes and isolated state for each experiment. Corrections remain accessible through HTML links without scripts or authentication.
Foundations: one Start/End completion
Actionability replaces old ClarifyingQuestions. Complete Start in a copy without changing Maf.Backlog; compare analysis 0 errors, action 1 error and clarified copy 0 errors. The final object represents an answer already obtained, not an automatic dialogue. Catalog and approval demos are complete. Catalog.Names defines the allowlist; components.json is no longer a current configuration exercise.
Advanced: manipulate complete programs
Memory makes correction and forgetting observable; RAG separates eligible sources, dense retrieval and citation; branches require an explicit terminal. Harness adds repetition, budget and deadline; MCP separates discovery and rights; streaming requires a terminal and persistence. Each page states actual model requirements. Scripted tests exercise boundaries without turning fixture success into model quality.
Consolidation and historical exercise disposition
| Mechanism | Scope |
|---|---|
| StatusTool / Task<bool> | Retired: no distributed status tool; study Found/NotFound/Refused/Error. |
| ResetAndList | Replaced by approval/refusal and readback; no reset at startup. |
| Rag_SourcesFormat | Exact citation and fabricated-citation rejection; no separate Start project. |
| SharedMemory_Facts / Reuse | Correction, isolation and coordinated forgetting in complete programs. |
| Otel_Tokens / ToolAndHttp | Nullable usage, native spans and separately reopened journal. |
Make results reviewable
A correction explains why a guard accepts or refuses. It grants no new permission and does not authorize removing validation. Dated receipt logs are not newly executed results on your machine. Do not redistribute retired historical solution ZIPs.