Leadership heard agents will improve themselves from past runs and cut Copilot credit spend. Makers want auto-accept. GRC wants a human gate. What’s actually coming, and how do we stop silent instruction drift in production agents?

Self-learning recommendations, maker-accepted refined copies, cost control — Steady Improve before unsupervised agent rewrites.

Can Copilot Studio agents ‘self-learn’ — and who approves the changes?

You’re right to flinch at “the agent will fix itself.” Unsupervised instruction drift is how last month’s safe bot becomes this month’s creative spender — with nobody sure which publish introduced the change.

Here’s what’s going on. Microsoft 365 Roadmap 570432 (Self-learning for Copilot Studio Agents, status In development; preview timing September CY2026 and GA October CY2026 on the roadmap — verify for your cloud) describes agents helping makers improve agents from completed runs. The platform identifies repeated tool sequences that could become workflows, to standardise behaviour, reduce errors, and control costs. Makers see recommendations with supporting evidence, expected impact, and proposed instruction/tool changes — then accept, configure, test, and publish a refined copy. That last phrase matters: the design keeps makers in control and preserves work that still needs model judgment. It is not “production rewrites itself overnight.” Broader Learn guidance on improving Copilot Studio projects still centres measurement, KPI reviews, prompt/instruction optimisation, and human iteration — self-learning fits that Improve loop, it doesn’t replace ALM or environment promotion. Until the feature is live in your tenant, treat 570432 as directional planning, not a button you can press today.

Steady before scale puts Pane 08 Cost & Value next to Pane 05 GRC: cost-saving recommendations that change tools are still change. Schofield’s Second Law (historically attributed) — have another copy — is why “refined copy” and test-before-publish beat in-place magic.

Options that actually work:

When 570432 lands, allow recommendations in sandbox only; require test evidence and a second publisher (or CoE review) before promoting the refined copy. Best Steady default for production agents.
Use recommendations as a cost clinic: prioritise repeated tool sequences that burn credits, convert the boring ones to workflows, leave judgment-heavy steps generative. Good CFO narrative with maker control.
Auto-accept every recommendation in production. Silent drift, weak audit story, and “the AI approved it” won’t help in an incident review.
Ban the feature entirely and never look at run evidence. You’ll keep paying for the same tool loops while competitors productise the Improve pane.

12 Panes action / comms: Tell makers: Recommendations are drafts — you still own the publish. Tell GRC: Refined copy = change record; keep the previous version until the new one proves out. Finance: Ask for expected credit impact on the recommendation card before you cheer the saving.

Benefits / risks: Standardising repeated sequences can cut errors and spend; preview/GA timing moves; bad recommendations accepted at speed amplify mistakes; “self-learning” in a slide deck will be oversold — stick to maker-accepted refined copies. Verify Learn and Message Centre when the feature appears in your environments.

I’d suggest this week: add “Improve / self-learning recommendations” to your AI register as In development, define who may promote a refined copy from sandbox, and pick one high-volume agent whose transcripts already show repeated tool spam — so you’re ready to trial when 570432 reaches you. You’ll get Steady Improve without unsupervised rewrites.

DIGITAL SKILL WORKS LTD / 12 PANES · You ask, we answer