How to Choose Microsoft Foundry Copilot Harness: Managed Agents, Hosted Agents, and Cost Checklist
The preview adds skills, tools, execution, and source-controlled definitions; sandbox compute and memory become metered on December 1
Key takeaways
Microsoft introduced the Copilot harness in Foundry Agent Service in public preview on October 28, 2026, while long-running resilience for hosted agents became generally available. The managed harness supplies an agent loop and sandbox with reusable skills, connected tools, file execution, bounded delegation, and source-controlled agent definitions. Existing agents do not switch automatically, and the harness cannot be changed after creation during the preview. Billing for harness runtime begins on December 1, 2026; model inference, tools, and connected services remain separately billed. This guide compares prompt agents, managed-harness agents, and hosted agents through control, recovery, idempotency, pricing, and verification requirements before a team commits to an architecture.
By ENHE AI · 2026-11-01 · AI-assisted drafting and editorial checks.
Direct answer: Choose by the layer you need to control, not by the longest feature list
Microsoft introduced the Copilot harness in Foundry Agent Service in public preview on October 28, 2026. A plain prompt agent fits tasks where instructions and supported tools are enough. Copilot harness fits complex work that needs a managed loop, skills, files, and delegation. Hosted agents fit teams that must bring their own framework, middleware, and runtime behavior.
The four capabilities in Copilot harness
The preview combines four building blocks in one managed execution system: reusable skills, connected tools, execution and delegation, and source-controlled agent definitions. Microsoft runs the loop and sandbox, while the team still supplies domain knowledge, procedures, tool authentication, approvals, and evaluation criteria.
Long-running recovery is not automatic idempotency
Long-running resilience for hosted agents is generally available and helps work that lasts hours or days resume after interruptions. Foundry restarts with the original input, but the application must save progress and intermediate results. If an agent creates reports, sends messages, or updates another system, it must check whether the action already happened before retrying.
Preview limits and migration boundaries
Existing agents remain unchanged. During the preview, the harness must be selected when the agent is created and cannot be switched afterward. Foundry manages CPU, memory, and idle settings. A team that needs to change the loop, dependencies, runtime behavior, or middleware should choose a hosted agent instead of assuming a managed-harness agent can be converted later.
Where the costs come from
Copilot harness runtime becomes consumption-priced on December 1, 2026. Before that date, harness runtime is free, but model inference, separately billed tools, and connected services still follow their own charges. Afterward, sandbox CPU and memory are also billed, with prices varying by region and agreement. A budget should track models, tools, CPU, memory, and external write operations separately.
A selection and verification checklist
Write down task input, permitted tools, human approval points, maximum duration, and external side effects. Use the same small task to compare the three approaches: custom-loop needs, interruption recovery, duplicate-write risk, log access, and cost attribution. Advance the least complex option that meets the evidence requirements.
Frequently asked questions
Will an existing Foundry agent automatically upgrade to Copilot harness?
No. Microsoft says existing agents remain unchanged, and the harness cannot be switched after creation during the preview.
Does long-running resilience prevent duplicate reports automatically?
Not automatically. The application still needs checkpoints and an existence check before retrying an external write.
Related tools and reading
These pages support the same task, workflow, or risk boundary.
- Microsoft Foundry agent-platform update
- Four-layer OpenShell runtime-safety checklist
- Four-step evidence checklist for agent autonomy
Conclusion
Use a plain prompt agent when instructions and a few supported tools are enough. Choose Copilot harness when the workflow needs skills, files, delegation, and a managed loop. Use a hosted agent only when the team must own the framework, middleware, or runtime behavior. Whichever route is chosen, validate it with one interruption-recovery exercise and one duplicate-write exercise.
What this means for everyday users
Existing agents remain unchanged. During the preview, the harness must be selected when the agent is created and cannot be switched afterward. Foundry manages CPU, memory, and idle settings. A team that needs to change the loop, dependencies, runtime behavior, or middleware should choose a hosted agent instead of assuming a managed-harness agent can be converted later.
Related tutorials
[[zh]][[/zh]][[en]]Review a public website[[/en]]
SearchScope SEO/GEO
[[zh]][[/zh]][[en]]Access and usage guide[[/en]]
AI Prompt Management System | 418 Bilingual Prompts for Writing, SEO and AI Creation
High-Frequency AI Prompts for Work, Learning, and Teaching guide
High-Frequency AI Prompts for Work, Learning, and Teaching
Related reading
Microsoft Foundry brings model choice, voice, and continuous optimization into one loop
Microsoft announced a set of Foundry Agent Service updates on September 24. Voice agents entered public preview with support for more than 80 languages and 140 locales, and resilience for long-running hosted agents also entered public preview. Tool search in Toolboxes and Agent-to-Agent support are generally available, while Routines can start agents on a schedule, after a delay, or from an event. Foundry also links production traces, issue insights, evaluation data, and agent optimization into a continuous improvement loop. Microsoft's token-reduction figures come from an internal evaluation on a public benchmark, so teams still need their own tool catalog, accents, telephony paths, permissions, and regression tests.
When Should an AI Agent Act Autonomously? A Scope, Know, Bound, Observe Checklist
An AI agent should receive more autonomy only when its actions can be verified, not because one demonstration looked successful. AWS describes a Trust in Depth framework organized around four lifecycle questions: Scope, Know, Bound, and Observe. This durable guide turns those questions into an operational checklist for a small team. Using an internal task-status update as the example, it identifies evidence required before launch, controls during execution, records after execution, and conditions that should stop the agent and request human review. Precise rules can sometimes be tested or formally checked, while tone, sensitive context, and high-impact decisions still require judgment. The objective is bounded, reviewable autonomy rather than either constant approval or unrestricted action.
How to Validate AI-Agent Runtime Safety: A Four-Layer OpenShell Evidence Checklist
NVIDIA launched the OpenShell Reference Architecture on October 21, 2026 for building, containing, observing, and enforcing safer agents across enterprise infrastructure and edge systems. This durable guide turns that architecture into a four-layer acceptance checklist rather than a product summary. For one low-risk agent task, teams verify build provenance, identity and tool boundaries, file and network isolation, observable action logs, policy denials, human approval, and rollback. Each layer requires evidence and a failure response. The goal is not to claim that one runtime makes an agent safe. It is to decide which actions can proceed automatically, which must stop for review, and how operators can reconstruct what happened after a failed or unexpected action.
How to Control Copilot Automations Cost: A Four-Layer Budget Checklist
GitHub expanded Copilot Automations on October 28, 2026 with external REST API triggers, 24 additional GitHub event triggers, 20 new tools, custom MCP tools, workflow chaining, and AI-credit limits for each run, day, week, or month. A per-run limit stops the current run when reached. Daily, weekly, and monthly limits prevent new runs until the relevant period resets. Automations without limits continue running under the existing behavior. This durable guide treats budgets as one control rather than a savings guarantee. It provides a sequence for bounding triggers and tools, setting four layers of limits, requiring approval for consequential actions, handling failure, and reviewing usage before expanding scope.
How to Validate Prompt Caching Savings for AI Agents on Bedrock in Four Steps
AWS published prompt-caching patterns for OpenAI models on Amazon Bedrock on October 26, 2026. An agent can reuse an identical prefix containing system instructions, tools, content, and history so a later request processes only the new portion. Bedrock supports implicit and explicit caching, while a cache key can share a reusable starting point across sessions or helper agents. AWS says cached prefixes have a minimum 30-minute time to live that is refreshed by successful reuse. This durable guide gives teams a four-step acceptance plan for stable-prefix design, cache-hit evidence, intentional invalidation after tool changes, latency measurement, and complete request cost instead of assuming every cache hit saves money.
When Should an AI Agent Use a Local Small Model or the Cloud? A Hybrid Routing Checklist
AWS published an Adaptive Agentic Intelligence Framework guide using Strands Agents on October 21, 2026. The reference routes routine status checks and local vision work to small models through Ollama, while complex analysis goes to Amazon Bedrock. This durable guide helps teams decide when an AI agent should stay local and when it should use a cloud model. It classifies tasks, marks data that may leave the device, establishes local and cloud quality baselines, defines timeouts and fallbacks, applies one safety policy across both paths, and measures actual routing before estimating cost. The example can run on an 8 GB Raspberry Pi, but that hardware detail is not evidence that every workload belongs on a small local model.