Both routes start with a working agent. The difference is how much of the surrounding product platform your team needs to connect.
CHAPTER I / THE ASSEMBLY ROUTE
The platform becomes your project.
Your agent works. Now connect everything your customers need around it.
01
Bring customers in.
Join your sign-in, customer roles, sessions and tool permissions.
02
Keep the work moving.
Connect databases, workflows, approval screens, schedules and channels.
03
Make it a business.
Attribute usage, enforce product entitlements and operate the connections.
Your team owns every join between the services.
CHAPTER II / THE CORE ROUTE
Your product stays the project.
Your agent works. Core connects it to the platform your customers need.
01
Sign them in. Keep their scope.
Core carries the user, customer and role through runs, data and tool calls.
02
Give the work a home.
Agents and workflows use shared controls. Managed databases and chat channels are planned.
03
Build the experience they buy.
Embed the UI, define the business rules and see usage by customer.
Your team owns the product your customers came for.
Roll out to the next customer.
Reuse scoped client identity and AI usage reporting. Managed database partitions are planned; your application supplies its business-data storage and access rules today.
Shared context for your integration.
Recover when a write times out.
The reviewed proposal, effect identity and uncertain outcome stay connected. Your operators get a recovery path, and your engineers have fewer failure-handling systems to invent.
Less recovery code to own.
Chat channels are planned.
The planned channel integrations would reuse published work, approvals and customer identity. Embedded web components and API integration are implemented for preview evaluation.
Web integration in preview.
THE DIFFERENCE, SIDE BY SIDE
What’s built in. What’s left to build.
Compare capabilities in the named products. Select a row to see what your team would build or connect. Engineering estimates are in the separate scenario below.
IncludedIncludedYesProvided by the named product
Not built inNot built inBuildYour app supplies the feature
Extra servicesExtra servicesAddConnect additional services
Provider-limitedProvider-limitedLimitLimited to the vendor’s models
Capability coverage for Core, Claude, OpenAI, Microsoft and AWS
Compare
CoreAgent platform
ClaudeManaged Agents
OpenAIAgents API
MicrosoftMSFoundry
AWSBedrock AgentCore
PreviewSigned user context
Not built inNot built inBuildCustom tool authorization
Not built inNot built inBuildRoles for custom functions
Not built inNot built inBuildTenant context in tools
Not built inNot built inBuildSession ownership in Instances
Customer roles in detail
Across runs and tool calls
CORE · PREVIEW
Your customer model, carried through.
Preview. Scoped client sessions and signed user context are implemented for evaluation.
App users, audiences and roles stay attached to runs and signed tool calls.
Onboard the next customer with a shared access model across tools, data and screens. Reduce repeated authorization work at every integration.
Claude Managed Agents
Not built in
Already provided
Session access and tool permissions.
Missing out of the box
Your SaaS customer roles carried into custom-tool authorization.
What your team supplies
Map users and audiences, then enforce business permissions in your tools.
OpenAI Agents API
Not built in
Already provided
Managed sessions and authenticated user IDs in ChatKit.
Missing out of the box
Your app’s customer-role policy across sessions, data and custom functions.
What your team supplies
Authenticate users, scope sessions and authorize each business tool.
Microsoft Foundry
Not built in
Already provided
Entra identity, RBAC and delegated user access.
Missing out of the box
Your SaaS customer context enforced across agent handoffs and tools.
What your team supplies
Propagate trusted tenant context and implement your app’s role rules.
AWS Bedrock AgentCore
Not built in
Already provided
IAM, identity and isolated sessions.
Missing out of the box
User-to-session ownership in shared-backend Runtime Instances deployments.
What your team supplies
Bind session IDs to users and apply your app’s customer-role rules.
PreviewReview and effect tracking
Not built inNot built inBuildCustom action review
Not built inNot built inBuildBusiness write recovery
Extra servicesExtra servicesAddConnect review and recovery
Not built inNot built inBuildBusiness review and recovery
Approvals & recovery in detail
Exact proposals and uncertain writes
CORE · PREVIEW
Approval follows the exact action.
Preview. Approvals and recovery are implemented; validate each business tool's outcome and permissions.
Version-bound review, repeat-safe effects and a recovery path when an outcome is uncertain.
Keep a timeout from becoming a blind second payment, email or update. Give operators a defined path to reconcile what happened.
Claude Managed Agents
Not built in
Already provided
Approval policies for agent and MCP tools.
Missing out of the box
Approval policies for application-executed custom tools.
What your team supplies
Implement custom-action review, effect records and timeout reconciliation.
OpenAI Agents API
Not built in
Already provided
Managed review controls and function calls handled by your code.
Missing out of the box
Your business proposal’s review and external-write reconciliation lifecycle.
What your team supplies
Bind approval to the proposal, then track and reconcile business effects.
Microsoft Foundry
Extra services
Already provided
Workflow approval gates and managed tool governance.
Missing as one connected feature
Proposal review and external-effect reconciliation as one business feature.
What your team supplies
Connect review state, effect keys and your reconciliation handlers.
AWS Bedrock AgentCore
Not built in
Already provided
Gateway policies govern whether a tool call is allowed.
Missing out of the box
Your exact-proposal review UI and uncertain-write resolution workflow.
What your team supplies
Build business review, effect records and reconciliation around policy checks.
PlannedManaged partitions (planned)
Extra servicesExtra servicesAddAdd an operational database
Extra servicesExtra servicesAddAdd an operational database
Extra servicesExtra servicesAddConnect a database service
Extra servicesExtra servicesAddConnect a database service
Customer database in detail
Partitions, migrations, export and deletion
CORE · PLANNED
Managed customer databases are planned.
Planned. Managed shared databases and audience partitions are planned. Use your application's database today.
Planned shared databases with audience partitions, schema migrations, export and deletion.
The planned service would manage customer partitions and their lifecycle. Today, connect your own database through business tools.
Claude Managed Agents
Extra services
Already provided
Session files and memory stores.
Missing as one connected feature
A managed operational database with customer partitions and schema lifecycle.
What your team supplies
Add a database and implement customer migrations, export and deletion.
OpenAI Agents API
Extra services
Already provided
Durable session state and sandbox files.
Missing as one connected feature
A managed customer-partitioned operational database and its lifecycle.
What your team supplies
Add a database and implement customer migrations, export and deletion.
Microsoft Foundry
Extra services
Already provided
Agent storage plus connected Azure data services.
Missing as one connected feature
One agent-platform database feature with customer partitions and lifecycle.
What your team supplies
Connect a database service and implement its customer data lifecycle.
AWS Bedrock AgentCore
Extra services
Already provided
AgentCore Memory and separate AWS database services.
Missing as one connected feature
A customer-partitioned operational database managed through the agent product.
What your team supplies
Connect a database and implement partition, schema and lifecycle operations.
Preview + plannedAI usage; billing planned
Not built inNot built inBuildCustomer usage and plans
Not built inNot built inBuildCustomer usage and plans
Not built inNot built inBuildCustomer usage and plans
Not built inNot built inBuildCustomer usage and plans
Usage & plans in detail
Customer attribution and entitlements
CORE · PREVIEW + PLANNED
Usage that belongs to your customers.
Preview + planned. AI usage estimates by customer are implemented. Platform credit accounting, paid billing, commercial tiers, and billing spend caps are planned.
AI usage estimates by customer are implemented. Platform credit accounting, commercial plan activation, billing, and billing spend caps are planned.
Use customer-attributed AI estimates when evaluating usage. Platform credit reporting and billing spend caps belong to the planned commercial model.
Claude Managed Agents
Not built in
Already provided
Provider usage reports and session budgets.
Missing out of the box
Your SaaS plans, customer usage attribution and product entitlements.
What your team supplies
Build the customer usage mapping and enforce your commercial plan rules.
OpenAI Agents API
Not built in
Already provided
Provider usage records and service billing.
Missing out of the box
Your SaaS plans, customer usage attribution and product entitlements.
What your team supplies
Build the customer usage mapping and enforce your commercial plan rules.
Microsoft Foundry
Not built in
Already provided
Azure metering, quotas and cost controls.
Missing out of the box
Your application’s customer plans and attributed agent-product usage.
What your team supplies
Map cloud consumption to customers and implement product entitlements.
AWS Bedrock AgentCore
Not built in
Already provided
Service metering, limits and agent tool-payment capabilities.
Missing out of the box
Your application’s customer plans and attributed agent-product usage.
What your team supplies
Map service usage to customers and enforce your product plan rules.
Preview + plannedShared identity and controls
Extra servicesExtra servicesAddAdd UI and chat adapters
Extra servicesExtra servicesAddConnect ChatKit and chat
Extra servicesExtra servicesAddConnect web and chat controls
Extra servicesExtra servicesAddAdd UI and channel routing
Web + chat in detail
Shared identity and action controls
CORE · PREVIEW + PLANNED
One platform wherever work starts.
Preview + planned. Embedded web UI, API, schedules, and events are implemented. Managed chat-channel installations are planned.
Embedded UI, API, schedules, and events use shared execution rules. Managed chat channels are planned.
The planned channels would reuse run state, review rules and customer identity. Start with the implemented web integration.
Claude Managed Agents
Extra services
Already provided
Agent APIs, sessions and scheduled execution.
Missing as one connected feature
Your embedded product UI and installed chat channels with shared app-role rules.
What your team supplies
Add the UI and channel adapters, then connect customer identity and reviews.
OpenAI Agents API
Extra services
Already provided
Agent APIs and the separate ChatKit UI integration.
Missing as one connected feature
A web-and-chat product layer with your shared customer and action policies.
What your team supplies
Connect ChatKit, chat adapters and your cross-channel business controls.
Microsoft Foundry
Extra services
Already provided
Teams and Copilot publishing.
Missing as one connected feature
Your customer-facing web app and other channels under one app policy.
What your team supplies
Integrate product UI, channel adapters and shared customer rules.
AWS Bedrock AgentCore
Extra services
Already provided
Runtime endpoints and tool integrations.
Missing as one connected feature
An embedded product UI and chat-channel layer with shared app controls.
What your team supplies
Add UI and channel routing, then connect customer access and review state.
Preview. Provider connections are implemented. Confirm the model and adapter enabled in your environment.
OpenAI, Claude, Bedrock and compatible endpoints alongside deterministic execution.
Choose a supported provider for each job while keeping your product’s access, workflow and review contracts. Model changes stay behind a common platform interface.
Claude Managed Agents
Provider-limited
Already provided
Claude models in the managed harness.
Unavailable in this harness
OpenAI or other model families through that same harness.
What your team supplies
Use another harness or integration to run other model families.
OpenAI Agents API
Provider-limited
Already provided
OpenAI models in the managed Codex harness.
Unavailable in this harness
Other model families through that same managed harness.
What your team supplies
Use another harness or the separate Agents SDK integration path.
Microsoft Foundry
Included
Already provided
Multiple supported model providers through Foundry and Agent Framework.
Your setup
Select and configure a supported provider.
AWS Bedrock AgentCore
Included
Already provided
Multiple models and frameworks through AgentCore.
Your setup
Select and configure a supported provider.
This table describes capability coverage. Your existing infrastructure, chosen integrations and product requirements determine the engineering work.
One hypothetical application, two implementation routes. Adjust the scope and team rates to explore the work involved. These are planning assumptions, not measured vendor comparisons.
01 BUILD VERSUS CONFIGURE
A customer request-review app.
One-time team labor in USD. Headline costs are rounded; the role breakdown below shows the calculation.
Both routes start here
An existing application and working agent, with identity and customer access policies in place.
Three ready business APIs, approved model credentials, and a team able to integrate them. Core access is assumed for its route.
The supplied Core UI meets the product’s needs. Domain logic, data selection and business authorization stay with the application.
The rollout being estimated
Review a request, draft a response, ask a reviewer, and submit the approved response. Add customer-scoped data, embedded web UI, one registered chat app, staging and production checks.
Includes planned capabilities: managed databases, chat channels and paid billing. This models an intended future rollout. See current availability.
New APIs, bespoke UI, data migration, model evaluations and formal compliance certification are outside both estimates. Subscriptions, usage and provider charges are separate.
BUILD · Assemble the surrounding platform≈ $85K–$150K712–1,260 h of total team effort
Build and verify the selected platform capabilities around an agent service.
CONFIGURE · Integrate Core≈ $1K–$2.1K9–18 h of total team effort
Configure the selected capabilities and complete application-specific release checks.
Illustrative task estimates, not measured benchmarks. Existing internal systems can reduce the assembly work; deselect capabilities you already have below.
Starts with Growth’s $599/month base, billed monthly. Add your expected usage. The model assumes Core provides platform deployment, operations and platform QA. Your app-specific release checks are estimated below. See platform pricing
ILLUSTRATIVE EFFORT DIFFERENCE17.4–31.3
person-weeks across development, DevOps, QA and security
ILLUSTRATIVE INITIAL COST DIFFERENCE≈ $83K–$150K
694–1,251 h valued at each role’s own rate
MAINTENANCE VALUE LESS CORE BUDGET≈ $2K–$6.7K
22–62 h of maintenance capacity, less your Core budget
ROLE / INITIAL EFFORTASSEMBLE WITH AGENT SERVICESCORE ROLLOUT & APP CHECKS
Development
Application code, configuration and integration
Assembly route420–752 h$52,500–$94,000
Core rollout4–8 h$500–$1,000
DevOps / SRE
Environments, deployment, monitoring and recovery
Assembly route92–160 h$11,500–$20,000
Core rollout1–2 h$125–$250
QA
Regression, customer boundaries and failure-path checks
Assembly route144–248 h$12,240–$21,080
Core rollout3–6 h$255–$510
Security review
Permissions, credentials and sensitive action review
Assembly route56–100 h$8,400–$15,000
Core rollout1–2 h$150–$300
Setup, rollout and cost assumptions
The 30–120 minute setup estimate is for one working end-to-end agent in an existing app. It includes agent publication, an existing sign-in bridge, the supplied UI, one ready tool endpoint and a smoke test. It is a planning target, not a measured onboarding benchmark.
The hypothetical rollout includes customer data, three ready business APIs, a registered chat app and release checks. Managed databases, channels and paid billing are planned. Its Core development allowance is 4–8 hours; DevOps, QA and security add their own hours in the role table. The model assumes the service supplies platform hosting, database operations, platform tests and platform security work.
Both wider-rollout columns start with an existing agent service. The assembly route covers the additional customer platform, its deployment and its verification. Each workstream assigns hours to individual roles; there is no QA or DevOps percentage surcharge on top of hours that already include them. Existing internal systems can reduce the work: deselect capabilities you already have.
Initial cost is the sum of each role’s hours × its rate. Rates are editable planning inputs. Savings use the assembly route’s low estimate minus Core’s high estimate at the low end, and the reverse at the high end. A person-week is 40 hours of effort, not a promised calendar week.
For all selected capabilities, monthly maintenance is 32–67 h for the assembly route and 5–10 h with Core. Monthly value uses each role’s maintenance hours and rate, then subtracts your Core budget. It can be negative. This values team capacity and does not imply payroll reductions.
New business APIs, bespoke product UI, data migration, model evaluations, procurement and formal compliance certification are outside both rollout estimates. Model tokens, other cloud services and channel fees depend on your workload. Add Core usage to the budget. These task estimates are not vendor benchmarks or a complete total-cost-of-ownership quote.
02 A SMALLER STARTING POINT
First-agent setup. A separate planning target.
30–120minutes of development effort
Publish an agent, connect your existing sign-in and tool, add the supplied UI, and run the complete journey.
$63–$250 of development effort at your selected development rate.
Starting conditions: an existing application, Core access, sign-in, an approved model account and one ready lookup API. Use the supplied UI to explain one support request, then smoke-test the journey. This task estimate is not a measured onboarding time.
BUILD THE NEXT CHAPTER
Put your engineering into your advantage.
Your customer experience. Your domain expertise. Core behind it.