Orbital / Microsoft Copilot harness in Foundry: a business evaluation guide

Microsoft Copilot harness in Foundry: a business evaluation guide.

Microsoft put its managed Copilot harness into Foundry public preview. Evaluate workload fit, controls, data, cost, portability and pilot evidence before adoption.

Managed agent runtime · Business evaluation

By Orbital Content Team · Published October 28, 2026

AI-assisted research and drafting with editorial review. Release, availability, feature, limit, pricing and data statements were checked against official Microsoft announcements and documentation on October 28, 2026. Microsoft product and performance claims are identified as vendor-reported. The evaluation framework and business implications are Orbital analysis. Orbital did not use the preview, deploy a Foundry agent, reproduce Microsoft's performance claims, inspect the harness sandbox, verify every model or region, test failure recovery, audit customer data flows, confirm account-specific pricing or review customer-specific terms.

Microsoft introduced the Microsoft Copilot harness in Foundry Agent Service as a public preview on October 28. The managed runtime coordinates the model, context, files, tools, reusable skills, sandbox and agent loop for multi-step work. Microsoft says the same harness family powers products across GitHub and Copilot and can work across multiple supported model families.

The business question is not whether a managed harness can produce an impressive demonstration. It is whether the business should hand one vendor more of the execution layer around an agent—and whether the resulting gains in speed and managed operations justify preview risk, control limits, cost, data flow and switching friction.

What Microsoft announced on October 28

Microsoft's official blog describes a multi-model harness as the orchestration layer around a model: instructions, context, execution environment and controls. The company says its harness supports more than 20 frontier models plus bring-your-own-key in the broader Copilot product family. The Foundry implementation has a narrower documented boundary: it is a public preview for new prompt agents, requires a supported reasoning-and-tool-calling model in an eligible region and does not currently support bring your own model.

Microsoft's Foundry announcement identifies four preview building blocks:

  • Reusable skills: packaged instructions, scripts and supporting resources that an agent can apply as a procedure.
  • Connected tools: supported Foundry tools and Model Context Protocol tools with configured authentication and approvals.
  • Managed execution and delegation: file and command work plus bounded delegation to built-in subagents.
  • Source-controlled definitions: agent configuration and supporting resources managed in the portal or with application code.

The initial preview must be selected when a new prompt agent is created. Microsoft says an existing agent cannot switch to the harness during this preview, and customers do not control the sandbox CPU, memory or idle settings.

Microsoft's October 28 Foundry post also says long-running resilience for hosted agents is generally available that week. The linked Microsoft Learn page was still labeled preview when checked on October 28. A buyer should resolve that documentation mismatch in writing for the exact region, account and workload before relying on an availability or service-level claim.

Separate documented facts from business implications

What the release supports—and what it does not establish
AreaDocumented statementPractical implicationNot established
AvailabilityThe Copilot harness in Foundry Agent Service is a public preview dated October 28, 2026.A team can run a bounded evaluation in an eligible Azure environment.A production SLA, production recommendation, feature stability or availability in every Foundry region.
Managed runtimeMicrosoft operates the agent loop, workspace and sandbox for a prompt agent.The buyer may avoid building and operating part of the orchestration layer.That the buyer can configure every runtime behavior, dependency, CPU, memory or lifecycle setting.
Tools and skillsThe preview supports reusable skills, selected tools, files, shell commands, web fetching and subagents.Complex procedures can be packaged without writing a custom agent framework.That every required tool, authentication method, approval policy or regional combination is supported.
Model choiceSupported reasoning-and-tool-calling models can be selected without rewriting the surrounding orchestration.A team may compare supported models behind one managed agent definition.Bring-your-own-model support in the Foundry preview, automatic optimal routing or equal behavior across models.
WorkspaceEach conversation has a separate workspace; state can survive normal sandbox idle, resume and replacement.Multi-step work can continue with conversation-scoped files and execution state.A durable business record, backup, system of record or guaranteed recovery of every incomplete action.
PriceHarness runtime is free before December 1, while inference and separately billed tools still incur charges; Microsoft publishes sandbox compute and memory rates beginning December 1.A pilot can measure runtime, inference, tool and labor usage before recurring harness billing begins.Total production cost, customer-specific discounts, future prices or savings compared with an existing architecture.
PerformanceMicrosoft reports that its Rust runtime reduces session overhead and time to first token and says the harness is used across Microsoft products.The implementation is mature enough to justify a measured comparison.Independent performance, reliability or cost results for the buyer's workload.

Choose the right agent architecture before selecting a product

Microsoft's own comparison is a useful starting point. There are three distinct choices:

Match ownership to the workflow
ApproachUse it whenWhat the business owns
Prompt agent without the Copilot harnessInstructions and a small set of supported tools are enough.Workflow definition, evaluation, connections, approvals and operations around the standard runtime.
Prompt agent with the Copilot harnessThe workflow needs skills, managed files and commands, subagents or more involved tool coordination without a custom loop.Business procedure, data boundaries, tool permissions, evaluation, human review and outcome accountability.
Hosted or custom agentThe team must change the framework, agent loop, middleware, dependencies, runtime policy or process model.More application and runtime logic, with greater flexibility and greater engineering burden.

Do not adopt the middle option merely because it is newer. If a deterministic automation, a standard prompt agent or an ordinary software integration can complete the task reliably, the added agent loop may increase cost and failure paths without improving the outcome. Use Orbital's business process audit checklist to document the current work before choosing an agent architecture.

Select one bounded workflow

A suitable first pilot has a clear trigger, approved inputs, a finite procedure, visible evidence, a reversible result and a manual fallback. Examples include preparing an internal research brief from approved public sources, reconciling a small set of non-sensitive records for review or producing a draft operating report that a named employee approves.

Reject or narrow a candidate when it:

  • moves money, changes access, deletes records, publishes publicly or sends customer messages without explicit approval;
  • requires the agent to invent policy or resolve legal, medical, financial, employment or other high-impact decisions;
  • depends on undocumented tribal knowledge or inconsistent source records;
  • cannot prove which system contains the authoritative result;
  • has no safe retry, duplicate prevention or rollback path;
  • uses data the pilot is not approved to process; or
  • cannot be compared with a current manual or deterministic baseline.

Define the candidate with Orbital's AI automation readiness checklist and estimate its complete economics with the AI automation ROI checklist.

Write the control contract before the agent definition

A managed harness operates execution infrastructure; it does not decide the business policy. Write a short control contract that records:

  1. Purpose: the one approved outcome and explicit non-goals.
  2. Owner: one business owner and one technical owner.
  3. Identity: the user, agent and connected-service identities involved.
  4. Data: permitted sources, fields, classifications, regions and retention.
  5. Tools: allowed tools, actions, destinations and credentials.
  6. Approval: actions that require a person and the evidence that person must see.
  7. Limits: maximum duration, steps, files, records, tokens, tool calls, spend and concurrency.
  8. Success: business-result tests, not only a fluent final response.
  9. Stop: prohibited actions, quality thresholds, security events and budget conditions that halt the run.
  10. Recovery: duplicate prevention, checkpoint, rollback and manual fallback.

Keep this contract outside the agent's prose instructions as an enforceable application and access policy where possible. A prompt saying “do not send without approval” is not a substitute for a tool permission that prevents sending.

Audit tools, skills, shell access and approvals

Microsoft documents five harness-native capabilities: file read, file write, shell, web fetching and subagents. New harness agents have all five enabled by default unless they are configured otherwise, subject to service and organization policy.

That default deserves an explicit review. Microsoft says shell is the broadest harness tool. Leaving shell enabled can permit file and network behavior even if separate file or web capabilities are disabled; those toggles are not a strict security boundary. Start from the minimum tool set the workflow needs, then verify the effective boundary in tests.

The preview also separates approval behavior:

  • Microsoft manages approval policy for harness-native tools during public preview.
  • The require_approval option applies to customer-configured MCP tools, not the native harness tools.
  • Connection-backed tools keep their own authentication, connection, approval and regional requirements.
  • A skill is a procedure, not a permission boundary. Its scripts and resources still operate through the tools and identities available to the agent.

For every tool, record the authentication method, scopes, record and action boundaries, approval setting, data sent, data returned, rate limits, cost, retry behavior and revocation procedure. Treat remote MCP servers and third-party agents as external trust boundaries. Validate their outputs as untrusted input rather than instructions.

Coordinate tool design with Orbital's business software integration checklist so systems of record, field contracts, retries and reconciliation are defined before an agent acts.

Map conversation state, durable records and external data flows

Microsoft says a harness conversation has its own workspace for files and execution state. Workspace state survives normal sandbox idle, resume and replacement, but the sandbox is not a customer-managed resource or durable output store. Important outputs should be saved to Foundry Files or another durable service.

That distinction matters. A completed agent response does not prove that a report, CRM record or project task was durably saved. After every consequential tool action, read the destination back and compare the saved record with the intended result.

Microsoft's Agent Service data documentation says prompts and outputs are not available to other customers or the named model providers and are not used to improve those providers' models. It also says stateful features can store messages and other content, stored data can be deleted by the customer, and processing location depends on the configured model and tools. External services have their own terms and data practices.

Create a data map for the pilot:

  • input source, owner, permitted purpose and classification;
  • conversation history, workspace files and generated outputs;
  • model endpoint and processing location;
  • tool requests, third-party destinations and subprocessors;
  • application, trace and security logs;
  • durable system of record and read-back verification;
  • retention, deletion, backup and legal hold; and
  • roles allowed to inspect prompts, files, tool arguments and results.

Do not assume the Foundry data commitment automatically covers a connected search, SaaS application, MCP server or customer-built tool. Review each leg of the flow.

Require traceable evidence and outcome tests

Foundry tracing can capture inputs, outputs, tool calls, results, token use, latency and retries in Azure Monitor Application Insights. That can support debugging and audit evidence, but it creates another sensitive data store. Microsoft recommends redacting personal data, secrets and credentials and applying production access controls and retention policies to traces.

Build an offline evaluation set before live use. Include:

  • normal, incomplete, conflicting and out-of-scope requests;
  • malicious or irrelevant instructions in documents and web content;
  • permission denial, expired credentials and unavailable tools;
  • timeouts, rate limits, partial writes and duplicate events;
  • an action that requires approval and one that is prohibited;
  • a subagent that returns weak or unsupported work;
  • a missing durable output and failed read-back; and
  • a restart or retry after a mid-task interruption.

Score the complete result: source quality, factual accuracy, tool choice, parameters, approval capture, saved destination state, duplicate prevention, latency, reviewer time, correction rate and total cost. Compare the harness agent with the current process and a simpler agent without the harness. The relevant question is not whether it can finish; it is whether it finishes reliably enough at an acceptable total cost and risk.

Model complete cost, not a free-preview headline

Microsoft says Copilot harness runtime use is free before December 1, 2026, but model inference and separately billed tools and connected services still apply. Starting December 1, the announcement lists sandbox compute at USD $0.0994 per vCPU-hour and memory at $0.0118 per GiB-hour, subject to region and agreement. Azure's pricing page also warns that model tokens, tool use, storage, search, connectors, licensed data and other services can be separate charges.

Measure cost per accepted business result:

  • model input, output, cached and reasoning tokens;
  • sandbox CPU, memory and idle behavior;
  • web search, code execution, file search, storage and network transfer;
  • MCP, Logic Apps, SharePoint, Fabric, search and other connected services;
  • Application Insights ingestion, storage and retention;
  • tool retries, failed runs, duplicate prevention and recovery;
  • human review, correction and exception handling;
  • engineering, security, privacy, legal and procurement work;
  • support, monitoring and regression evaluation; and
  • the manual process retained for outages or unsafe cases.

Run the same measured workload through the current process, a standard prompt agent and the harness candidate. A managed runtime may reduce engineering ownership while increasing platform usage or switching cost. Only a complete comparison can show the trade.

Ask Microsoft and implementers for proof

Before production procurement, request written answers for the exact subscription, region and architecture:

  • public-preview terms, SLA, support path and transition plan;
  • supported models, tools, authentication methods and regional combinations;
  • effective controls for shell, web, files, subagents and remote MCP;
  • approval behavior for native and customer-configured tools;
  • sandbox isolation, patching, replacement and evidence available to customers;
  • conversation, workspace, file, trace and backup retention;
  • data-processing locations and terms for every model and tool;
  • customer-managed key and private-network support for required features;
  • quotas, concurrency, throttling and capacity commitments;
  • billing meters, idle rules, discounts and cost-export fields;
  • agent definition, skill, trace and data export for migration; and
  • ownership and response procedure for incidents or harmful actions.

Ask the implementer for a live demonstration using a synthetic failure case, not only a happy path. Require a permission denial, malicious tool response, duplicate event, timeout, stopped run, recovery, human approval and durable read-back.

A four-week bounded pilot

  1. Week 1 — define and baseline. Choose one internal, reversible workflow. Document purpose, owner, current time and error rate, data boundary, approval points, cost and stop conditions. Confirm preview terms and supported region.
  2. Week 2 — build the minimum agent. Create a new non-production agent. Disable unnecessary native capabilities, attach only approved tools and skills, use narrow identities and configure durable output and traces with minimization.
  3. Week 3 — test offline and in shadow mode. Run fixed normal, edge, attack and failure cases. Then process permitted live work without changing the real system. Compare the harness, simpler agent and existing process.
  4. Week 4 — release one reversible step. Allow one bounded internal action behind human approval. Measure accepted outcomes, corrections, review minutes, latency, retries, tool failures, trace completeness and total cost. Exercise the stop and manual fallback.

Stop rather than expand if the agent crosses its authority, cannot explain or evidence tool actions, loses durable output, duplicates work, leaks sensitive data into traces, creates an unacceptable review burden or costs more than the measured benefit.

Orbital's business take

The Copilot harness announcement is important because the model is becoming one replaceable component inside a larger operating system for agents. For buyers, that can be useful: managed context, tools, skills, files, sandboxing and subagents may reduce the amount of orchestration infrastructure a team has to own.

The trade is deeper platform dependency. The preview controls the runtime, supports only selected regions and models, enables broad native capabilities by default, limits some request settings and does not support BYOM. “Multi-model” should not be translated into “model-independent,” and “managed” should not be translated into “safe by default.”

A service business should evaluate the harness only after it can name one measurable workflow, enforce tool permissions outside the prompt, verify the saved business result and operate a manual fallback. If the managed runtime improves accepted outcomes or reduces engineering and review cost in a bounded pilot, there is evidence for a larger evaluation. If not, a simpler automation may be the better architecture.

Explore Orbital's AI automation services, tech stack consulting, connected-business planning, business consulting and service-business planning. A 60-minute consultation can turn one candidate workflow into a scoped evaluation brief; implementation, Azure usage, model and tool fees, software subscriptions, security, privacy, legal review and ongoing support are separately scoped.

Sources and editorial limits

Primary Microsoft sources were checked October 28, 2026. Microsoft dated both launch posts October 28. The Copilot harness is reported as a public preview; Microsoft's preview documentation says preview features have no service-level agreement and are not recommended for production workloads. Microsoft marketing statements about scale, performance and product use were not independently reproduced. The Foundry announcement says long-running hosted-agent resilience is generally available that week, while the linked Learn page was still labeled preview when checked; buyers should verify the applicable status and terms. Orbital did not use the preview, deploy an agent, test sandbox isolation, measure performance or cost, verify every model, tool or region, inspect Microsoft systems, audit a customer environment or review customer-specific contracts. This article is business analysis, not legal, privacy, security, employment, financial or other professional advice, and it does not guarantee accuracy, availability, savings, compliance or business results.

Primary source links

Your next chapter.
Let’s give it lift.

Start a conversation