Capability baseline and update record

Teaching baseline: summer 2026-era agent workflows. Documentation verified September 8, 2026. Current pages can include later releases; this package does not claim they establish historical availability in June, July or August.

The course teaches multimodal evidence, repository exploration and edits, executing code/tests, structured data work, tool use, reusable instructions, connectors/MCP, browser assistance where enabled, and bounded multistep execution. A model's reasoning ability, the client's tools, and the user's account permissions are separate prerequisites. There is no promise that every learner has every connector or that native Discord/Google access exists.

The project recommendation is Qwen3.8-27B, released in summer 2026. The small-model guide, checked September 10, 2026, records the exact model, source links, hosted/local options, budget and required client compatibility check. No live model evaluation is implied by documentation review. Keep model choice separate from the Codex/OpenCode/Claude Code client and hosting provider.

Current official sources:

https://learn.chatgpt.com/guides/best-practices — task scope, reasoning effort, repository guidance and verification.
https://learn.chatgpt.com/docs/extend/mcp?surface=cli — tool/context connections and configuration.
https://learn.chatgpt.com/docs/build-skills — reusable instructions, resources and optional scripts.
https://learn.chatgpt.com/use-cases — business and engineering examples.

Instructor baseline card, completed before delivery:

Model/provider and exact ID visible in this cohort:
Client version and host:
Allowed tools actually available:
Approved data sources and capabilities:
Tasks executed on the eight-case corpus:
Observed limits and fallback:
Date checked:

Do not use 2024/2025 chatbot limitations to restrict current agents to drafting. Within authorized local work they can inspect files, implement changes, run tests, and iterate to a verified result. Extra reasoning, tools and extra agents have costs; choose them by task outcomes. Independent review reduces some errors, but reviewers can share blind spots. Measurements govern adoption.

The synthetic A/B outputs in fixtures are for learning scoring. They are not performance evidence for any model. Record actual model runs separately before making a procurement or deployment decision.