Pick the path before you pick the demo.
Microsoft 365 Copilot agents now ship on two engineering paths: declarative agents that ride Copilot's orchestrator and foundation models, and custom engine agents that bring external orchestration and hosting: with a Copilot Control System framing how security, lifecycle, and analytics gate rollout. Microsoft's documentation treats the fork as a compliance and cost decision, not a branding choice.
Declarative agents add custom instructions, knowledge connectors, and actions while inheriting Microsoft 365 security, compliance, and Responsible AI requirements. Custom engine agents suit complex workflows, proprietary models, and proactive messaging: but Microsoft states operators must ensure their own compliance, RAI practices, and security measures, typically on Azure or similar cloud hosting at additional cost.
This Field Note maps Microsoft's published agent anatomy and Control System pillars to operator weeks. It draws on the agents overview and Copilot Control System documentation in Sources. Named customer agent deployment counts are UNKNOWN.
Job the system was hired to do
Extend Copilot with domain-specific knowledge and actions inside Teams, Outlook, Word, and Copilot Chat: without rebuilding permission boundaries Microsoft Graph already enforces for declarative agents, and with explicit governance when custom engines call external APIs.
Buyers hire agents to automate multi-step workflows: IT helpdesk @mentions in Teams, sales lead creation in Dynamics or Salesforce, document summarization invoked from SharePoint. Microsoft's agents overview lists core components: knowledge, actions, orchestrator, foundation models, user experience layer: and assigns hosting responsibility based on agent type.
Declarative agents were hired when the job fits Copilot's orchestrator: focused scenarios, user-initiated interactions, integration with M365 data via Graph and Copilot connectors, no additional model hosting. Custom engine agents were hired when the job needs custom orchestration, domain-specific or multimodal models, proactive triggers without user input, agent-to-agent collaboration, or channels outside M365.
The Copilot Control System was hired to govern whichever path wins: secure data Copilot and agents reference, manage licensing and agent lifecycle, and measure adoption and ROI through Copilot Analytics.
First week
Read Microsoft's decision table comparing declarative and custom engine agents on use case, sharing, customization, proactive interactions, channels, setup complexity, engine hosting, deployment options, and compliance inheritance. Week one deliverable is a written choice memo, not code.
For declarative candidates, scaffold with Microsoft 365 Copilot low-code Agent Builder or pro-code Visual Studio Code plus Microsoft 365 Agents Toolkit. Microsoft documents Work IQ Dev Tools (preview): wiqd: for terminal-based lifecycle: scaffold, validate, provision, package, publish, monitor. Week one often produces a single-agent manifest with custom instructions and one SharePoint or Copilot connector knowledge source.
For custom engine candidates, inventory existing conversational assistants outside Copilot. Microsoft recommends custom engine when integrating prior bots with M365 channels or when Copilot Studio or Semantic Kernel orchestration must call multiple credit-check APIs with business rules Microsoft's declarative orchestrator can't express.
Parallel week-one work on the Copilot Control System security pillar: review data security, AI security, and compliance and privacy documentation linked from the overview. Management controls live in Microsoft 365 admin center, Power Platform admin center, and Copilot Studio: licensing and metering, agent lifecycle, customization settings. Measurement and reporting routes through Copilot Analytics for readiness, adoption, productivity impact, and ROI.
Pilot cohort selection stays narrow. Microsoft notes declarative agents are designed for individual use; custom engine agents support group collaboration in Teams channels. Week-one pilots pick one channel: often IT Teams chat: with Purview logging assumptions inherited from base Copilot deployment.
Cost review belongs in week one. Microsoft links separate cost considerations for Copilot extensibility; custom engine hosting on Azure adds metered spend beyond Copilot seats. Exact pilot dollar burn is contract-specific and UNKNOWN here.
What broke or was routed around
Teams that built custom engine agents for Graph-scoped Q&A broke compliance inheritance. Microsoft states declarative agents adhere to M365 RAI and security standards automatically; custom engines shift that burden to the builder. Operators routed simple FAQ and document Q&A back to declarative agents after legal flagged missing inheritance attestations.
Proactive messaging requests broke declarative assumptions. Declarative agents rely on user-initiated interactions; proactive workflows require custom engine agents with programmatic triggers. Product teams that wanted autonomous ticket escalation without user @mention hit the declarative ceiling and either scoped down to user-initiated flows or opened a custom engine workstream with longer security review.
Agent sprawl without Control System lifecycle management broke admin center visibility. Microsoft documents agent lifecycle controls in the management pillar; agents published from Copilot Studio to M365 channels without inventory landed on help-desk tickets asking which bot had access to which connector. Operators routed new publishes through a named approver list and Copilot Analytics readiness checks.
Overbroad connector actions broke least-privilege stories. Declarative custom actions integrate external APIs in real time; credentials and scopes are human-configured. Security teams narrowed action sets to read-only lookups until DLP and Purview agent policies matched data classification.
Custom model hosting costs broke pilot budgets. Microsoft documents additional Azure hosting for external orchestrators and models. Teams that estimated Copilot seat price only: without Azure OpenAI or Container Apps lines: paused custom engine work until finance signed a hosting envelope.
Multi-agent collaboration broke support models. Custom engine agents support agent-to-agent communication; declarative agents don't document the same delegation pattern. Operators routed multi-agent designs behind a single custom engine facade rather than many declarative agents chaining silently.
EU and residency subprocessors complicated custom engine model choice where declarative Copilot residency commitments differ from bring-your-own-model paths. Exact EU tenant delays are UNKNOWN in public Microsoft metrics.
Sales Agent and other Microsoft prebuilt agents broke uniform governance assumptions. Microsoft cites Sales Agent automating lead management across Dynamics or Salesforce as an example of specialized assistants. Operators treated prebuilts as vendor-managed for updates but still routed enablement through the same Control System lifecycle gates as custom builds: who gets the agent, which CRM credentials attach, and whether Analytics attributes ROI to prebuilt versus custom workstreams.
Low-code versus pro-code tooling splits broke staffing plans. Microsoft documents Agent Builder for declarative agents and Copilot Studio or Agents Toolkit for both paths. Teams that assigned agents only to pro-code squads delayed declarative pilots; teams that banned pro-code could not implement custom engine orchestration when declarative limits appeared. Operators routed a paired team model: declarative agents owned by M365 admins with security review, custom engines owned by application teams with Azure cost center billing.
What a person still does
Writes the declarative-vs-custom decision memo using Microsoft's flow chart criteria: orchestration complexity, model choice, proactive needs, channel scope, compliance inheritance.
Configures custom instructions and knowledge sources for declarative agents; designs orchestration graphs and model endpoints for custom engines. Microsoft supplies tooling; domain policy language is human-authored.
Owns Copilot Control System management controls: who receives agent licenses, which agents appear in which M365 apps, connector credential rotation.
Reads Copilot Analytics exports for adoption and productivity claims before scaling cohorts. Microsoft documents measurement pillars; interpretation and executive narrative remain operator work.
Reviews agent vendor terms when enabling third-party or partner agents. Microsoft warns agents carry their own privacy statements.
Handles exceptions when agents mis-route or hallucinate citations: same human-send discipline as base Copilot drafts on high-liability queues.
Cost / time in operator units
Microsoft documents Copilot seat licensing plus potential Azure hosting for custom engine agents and Copilot Studio consumption. Declarative agents avoid additional engine hosting per Microsoft's comparison table; custom engines add cloud meter lines operators must forecast.
Week-scale declarative agent from scaffold to Teams pilot is often one to three engineering weeks for a single connector scope, based on operator reports; Microsoft doesn't publish universal timelines. Custom engine agents with existing bot reuse may shorten integration; greenfield orchestration often runs longer.
Admin time concentrates in Copilot Control System pillars: security reviews, agent lifecycle approvals, Analytics dashboard configuration. Copilot Analytics ROI numbers require customer-defined baselines; Microsoft doesn't ship median hours saved per agent interaction in the overview docs.
Failure cost is permission leakage or non-compliant proactive messaging: reputational and legal, not a line item in agent SKU pricing. Routing declarative agents before permission remediation repeats base Copilot oversharing patterns at the connector edge.
What they would do next time
Default to declarative agents when the workflow is user-initiated, M365-scoped, and fits Copilot's orchestrator: inherit compliance instead of rebuilding it on Azure.
Stand up Copilot Control System inventory before publishing agent two: lifecycle owner, Analytics baseline, connector scope documentation.
Reserve custom engine agents for documented triggers: proactive automation, proprietary models, multi-system orchestration with business rules, or non-M365 channels.
Run Microsoft's cost considerations doc before custom engine approval; seat price alone is not total cost of ownership.
Pilot one channel with Purview logging enabled; agent prompts and responses follow the same retention paths as Copilot chat when policies require.
Use wiqd or Agents Toolkit validation steps in CI before production publish: Microsoft documents validate and package in the declarative lifecycle.
Publish internal guidance on @mention versus proactive triggers so product teams don't spec undeclarable behavior into declarative agents.
Re-read the comparison table when vendors sell custom engine complexity for Graph-only Q&A: declarative path is often sufficient and faster to govern.