You can build a declarative Copilot Studio agent in an afternoon. Getting it past security review is the part that stalls. The instruction block is the easy 20 percent. The other 80 percent is undocumented. You scope the knowledge source so security signs off, tell your DPO what the agent does with data, forecast what a run costs under the metered layer, then prove the agent refuses when it should. This pack ships that 80 percent for six agents. Each one arrives finished for governance, not for a demo. You get its paste-ready instruction block, its knowledge-source scoping, a governance note for your DPO, a labeled-assumptions cost estimate under the metered layer, then a test script that proves the refusals fire. The six agents are Meeting-to-Action, Policy and SOP Answer, Vendor and Procurement Intake, New-Hire Onboarding Buddy, Weekly Status Digest, and Inbox Triage and Draft. None of them is safety-critical, and that is deliberate. Version 1.1 adds the half that only shows up later. Six agents shipped is not six agents governed: three months on, someone asks who approved this one, what it reaches, whose budget the metered line lands on, and who turns it off. So the pack now carries an operating layer. A 23-field agent register workbook with live formulas is its spine, and beside it sit five standards and a note: naming and ownership, an approval-gate RACI with eight gates and exactly one accountable name each, a nine-step decommissioning SOP, the refusal fence written out as a clause you paste into every agent, and one page on how all of this sits next to an agent control plane. Every version is a free update to anyone who has already bought; you re-download from your Gumroad library. Why the honesty matters here: the six instruction blocks overlap by design with the free awesome-copilot-studio-agents repo and Microsoft's own template library. The instruction text is the part everyone already gives away. What no free source ships is the refusal clause proven by an included test, the hand-to-DPO governance note, and the labeled-assumption cost estimate. That is the 80 percent, and that is what the price buys. It all arrives as a single download, one .zip holding the guide, the folders, the workbooks, a changelog, and the license, plus a read-me with a real contact address (mathieu@kesslernity.com) so you know who stands behind it. What's inside, and what each file actually is, is listed below. What's inside Six paste-ready Copilot Studio instruction blocks, 1,500 to 2,000 characters each, delivered as plain .txt in an instruction-blocks/ folder so they paste into the Instructions field without markdown or smart-quote corruption. Each has [BRACKETED] fill-in slots for your org names, source paths, owner, and refusal boundaries, and runs the same anatomy: role, scope and grounding with "Allow ungrounded responses" set OFF, refusal rules, a "what you never do" clause, a first-turn AI-disclosure line, and escalation to a named human. The Vendor block carries a mandatory-field hard stop; the Inbox block carries a never-auto-send rule that also appears in its test. The same six agents again as import-ready declarative agent packages, in an import-ready-packages/ folder, one .zip per agent, so you can upload instead of paste. An IMPORT-GUIDE PDF covers the upload path, what a package carries across tenants and what it does not, and what to check on the imported agent before you publish it. Pasting the block and importing the package produce the same agent; pick whichever your environment allows. An agent operating register workbook, four tabs, one row per agent across 23 fields: identity, the three named owners plus the co-owner who keeps the agent alive when one person leaves, what it reaches and the one line justifying that scope, its run mode, its approval and review dates, the metered cost line and the budget that carries it, and the written condition that is supposed to end its life. Three columns are live formulas, so the sheet tells you what each agent actually runs as, when its next review falls due, and whether that review is OK, DUE SOON or OVERDUE. Tab two is a connector and MCP pre-flight sheet, tab three logs decommissions, and six rows arrive pre-filled with the guide's worked example, reconciling with the cost estimator, so you can see the shape before you clear them. Six standards as standalone Markdown files in an operating-layer/ folder, so you can hand one to a platform owner or paste it into your governance wiki without sending the whole guide: the register field standard, argued field by field with what breaks if you leave one blank; the naming and ownership standard, including the validator regex and why a retired name never comes back; the approval-gate RACI, eight gates and seven roles with exactly one accountable name per gate; the nine-step decommissioning SOP, from the trigger firing to the row closing; the refusal fence, a paste-in clause in three layers with nine tests that prove it fires; and a one-page note on running this next to an agent control plane, including the quarterly reconciliation and the three failure classes it surfaces. A tickable checklists workbook with two tabs and a Done/Blocked dropdown on every row, so you can track each agent as you ship it. Tab one is a 10-item Governance and DLP checklist with a done-when condition on each item: it covers the trap that a Purview DLP policy scoped to Copilot does not cover custom Copilot Studio agents, least-privilege scoping, the maker-credential oversharing risk on triggered agents, RCD containment, the Article 50 disclosure line, the Flex Routing DPO flag, and the one-paragraph "what to tell your DPO" note. Usable against any agent, not only these six. Both checklists are also printed as readable pages inside the guide PDF. Tab two of that same workbook is a 9-step deployment runbook that ties the pack together, from confirming publish rights to setting a monthly review date, with a done-when condition per step and a target of under an afternoon per agent. A 2-tab cost estimator with live formulas and a zero-rating toggle. The Read Me tab states the published $0.01-per-credit rate and labels the credits-per-run figures as assumptions. The Calculator tab holds one row per agent with visible formulas, a Y/N metered switch that zeroes a row when internal use is zero-rated, and a low, base, and high band. It ships pre-populated with the Aldermont example and recomputes when you overwrite the inputs. A 7-tab test-script workbook: a Read Me plus one tab per agent. Each tab is a runnable log with columns for the prompt, the expected behavior, the expected refusal, an Actual column, a Pass/Fail cell, and Notes, pre-filled with each agent's acceptance and refusal cases so you can prove the refusals before you publish. A 55-page typeset PDF in eleven sections. Six of them walk an agent through the same seven-part structure (the 5-minute brief, the instruction block, topics and triggers, knowledge-source scoping, the governance note, the cost estimate, the test script) so you learn the shape once and repeat it. The last three are the prose behind the operating layer: the register and how to fill it, retiring an agent, and the refusal fence. Both checklists and all six standards print as readable appended pages, and it closes on a License and Contact page. FAQ Q1: What exactly do I get? One download, a single .zip. Inside: a 55-page guide PDF; six paste-ready instruction blocks as plain .txt; the same six agents as import-ready packages with their own import guide; four workbooks (the agent operating register, a cost estimator, a test-script log, and a tickable governance-and-deployment checklists workbook); the six operating-layer standards as standalone Markdown; plus a CHANGELOG, a LICENSE file, and a read-me with my contact address. Each deliverable is usable on its own without reading the prose. You download it the moment you buy, and every later version is free from your library. Q2: Will it work in my setup? You need a Copilot Studio environment, an M365 tenant with a DLP policy, and publish rights. These are declarative agents grounded on knowledge sources. No custom connector, Power Automate flow, plugin, or code is required, and none is included. Q3: How long does it take to ship an agent? The target is under an afternoon per agent if you follow the 9-step runbook: paste the block, fill the brackets, scope the source, run the governance checklist, review the cost line, run the test script, publish. Slower on your first agent, faster by the third. Q4: Isn't this the same as your free awesome-copilot-studio-agents repo, or Microsoft's templates? Honestly, the instruction blocks overlap with both, on purpose. The repo and Microsoft's library give you the instruction block, and that is the part everyone gives away for free. What neither ships is the omitted 80 percent: a refusal clause proven by an included test, a hand-to-DPO governance note, and a labeled-assumption cost estimate. That is what $49 buys. If you only want the instruction text, take it from the free repo with my blessing. Q5: I already have Agent 365, or my account team is selling it to me. Do I need this? They solve different halves, and if you have the control plane, keep it and run this beside it. Agent 365 reached general availability on May 1 2026, and Entra Agent ID gives every agent a first-class directory identity you can inventory, grant, monitor and revoke. That closes a real gap. What it governs is agents once they exist. Whether an agent should exist at all, who carries the consequence when it is wrong, the one line justifying the access it was granted, whose budget the metered line lands on, and the written condition that ends its life are decisions taken before the directory ever sees the agent, and none of them is a directory fact. The pack includes a one-page note on running both, with the quarterly reconciliation between the inventory and the register and the three failure classes it surfaces. On commercials, stated plainly because the numbers move: Agent 365 appears inside the E7 bundle at a $99 per user per month list price alongside E5, Microsoft 365 Copilot and the Entra Suite, agent runtime still bills through the metered credit layer rather than being included in a governance seat, and the standalone price is not stated on Microsoft Learn, so this page does not present a figure for it. Confirm your own packaging with your Microsoft account team. Q6: Is this just AI-generated fluff? No, and here is how to check. I drafted it against a research brief with named, dated sources, then it passed the factory's quality gates: a deterministic slop scan that returned zero errors, a build check that opens the PDF and recomputes both workbooks in a real spreadsheet engine, then a bullet-to-file check confirming every claim on this page maps to a shipped file. A human reviewed the voice and the numbers. Where Microsoft publishes a rate but not a per-task credit count, the cost figures say so and are labeled assumptions. Q7: Won't six metered agents blow up my Copilot bill? Usually not, and the estimator shows you why. Microsoft added a metered layer in June 2026 on top of the flat $30-per-user seat; the seat price did not change. For internal, interactive use by people who already hold an M365 Copilot license, agent use is zero-rated, so most of these agents cost effectively $0 on top of the seat. Metered cost shows up where the workbook flags it: scheduled or triggered runs, external or unlicensed users, tenant-graph grounding, the Cowork layer. The estimator ships with the Aldermont example ($1,034 per month at base-case for the metered cases, a low of $517 to a high of $2,068) so you budget a range and review monthly. Q8: Does this help with the EU AI Act and my DPO? It gives you considerations to hand them, not a legal opinion. The governance note treats these agents as limited-risk productivity tools, which is what they are while you stay inside Vol. 1's scope; the moment an agent makes an HR, credit, or biometric decision it can flip to high-risk, which is exactly what this volume excludes. The checklist includes the Article 50 transparency duty in force from August 2 2026 as a first-turn "you are talking to an AI" line, and it flags Flex Routing (default-on since April 17 2026, where LLM inference may leave the EU boundary) for your DPO. Your DPO and counsel own the assessment. Q9: Will it break when Microsoft changes the Copilot Studio UI? The instruction blocks are plain text you paste into the Instructions field; they do not depend on where a button sits, so a UI refresh does not touch them. What can date is a specific figure: the $0.01 credit rate, the 8,000-character instruction ceiling, a menu path in the runbook. Each is labeled with its "as of" date, mostly June 2026, so you can re-verify against Microsoft's current docs. When Microsoft moves a control, the runbook step still tells you what to accomplish. Dated, not frozen. Q10: What is the license, and can I get a refund? Licensed, not sold: one commercial license per buyer, for use inside your own organization, not for resale or redistribution as a competing product. Full License and Terms apply (ref KESS-LIC-2026-001), on the Kesslernity license page. These are digital goods; if the pack is not what this page describes, write to me and I will make it right. Licensed, not sold. Full License and Terms apply (ref KESS-LIC-2026-001); see the Kesslernity License & Terms. By purchasing you agree to the version in force on your purchase date. Kesslernity is an independent publisher. This pack is not affiliated with or endorsed by Microsoft. The governance and cost material is a set of considerations for your own DPO and counsel to approve, not professional or legal advice. Questions: mathieu@kesslernity.com