Home › Agent Harness › Architecture map
AI · Agent Harness
Agent harness architecture map: what fills each concern, on every stack
An agent harness is the same set of concerns whatever you build it on: a loop, a model door, tools, state, waits, identity, records, configuration. What changes between Azure, Databricks, AWS, GCP and an open-source stack is only the product that fills each one. This page puts all of them on one view, with a verdict on what to pick.
Written 20 November 2026 from the open stack files, one per platform. Open Stack is the only column that has run end to end. Azure and AWS are decided in writing and not yet built. Databricks is part built. GCP has no stack file at all: that column is a proposal, shaded grey, to be checked before anyone relies on it. Product names and features change; check current documentation.
The map
The table
One row per concern. Read across to see what each stack uses. Read the last column for what to select.
Cloud service the platform sells itOpen source a product you can run anywhereBuilt by us a few lines of our own, identical on every stack
built and run: it has run in a real builddecided, not built: chosen in a stack file, no running agent yetproposed: written for this page, no stack file
| Concern family | Concern | Azure | Databricks | AWS | GCP (proposed) | Open stack | Verdict: what to select |
|---|---|---|---|---|---|---|---|
| The loop and where it runs | Control loopWho owns the step budget, the no-progress check and the cost ceiling. | Built by us Our bounded loop in a containerdecided, not built | Built by us Our bounded loop, wrapped in the Agent Framework interface so it can be traced and loggeddecided, not built | Built by us Our bounded loop in a container on ECS Fargatedecided, not built | Built by us Our bounded loop on Cloud Runproposed, no stack file | Built by us Our bounded loopbuilt and run | Build it, the same on every stack. A loop you did not write cannot be made to stop where your specification says. Cloud agent runtimes can wrap it; they should not own it. |
| HostingWhere the agent, its tool servers and the review desk run. | Cloud service Container Appsdecided, not built | Cloud service Lakeflow job for batch work; Databricks App for an interactive deskdecided, not built | Cloud service ECS with Fargate (App Runner is closed to new customers; Lambda cannot hold a long wait)decided, not built | Cloud service Cloud Runproposed, no stack file | Open source Containers from a compose filebuilt and run | Use the container platform of the cloud you are already on. On Databricks a job fits batch investigations. | |
| Models | Model accessThe one door every model call goes through: keys, limits, usage records. | Cloud service API Management as the AI gatewaydecided, not built | Cloud service Foundation Model APIs; AI Gateway on top where a serving endpoint is allowedbuilt and run | Cloud service AgentCore Gateway in front of Bedrock or other providersdecided, not built | Cloud service Apigee or API Gateway in front of Vertex AIproposed, no stack file | Open source LiteLLM proxybuilt and run | Use the gateway of the cloud where the models and the data sit. Use LiteLLM when no enterprise gateway exists. One enforcement point per path. |
| Admission and limitsHow many calls run at once, who waits, who is turned away. | Built by us Semaphores in the harnessdecided, not built | Built by us Semaphores in the harnessdecided, not built | Built by us Semaphores in the harnessdecided, not built | Built by us Semaphores in the harnessproposed, no stack file | Built by us Semaphores in the harnessbuilt and run | Build it. A few lines of your own beat any product, and it must see the unit of work. | |
| Cost accountingSpend added up per task, with a ceiling that stops work. | Built by us Per-task meterdecided, not built | Built by us Per-task meterdecided, not built | Built by us Per-task meterdecided, not built | Built by us Per-task meterproposed, no stack file | Built by us Per-task meterbuilt and run | Build it. Cost per successful task needs your definition of successful. Cloud billing tables are a cross-check. | |
| Tools and knowledge | Tool runtimeHow a tool call is run, under whose authority. | Open source MCP clientdecided, not built | Cloud service Unity Catalog functions: typed, permissioned, no free SQL from the modelbuilt and run | Open source MCP clientdecided, not built | Open source MCP clientproposed, no stack file | Open source MCP clientbuilt and run | MCP everywhere. On Databricks, register the data tools as Unity Catalog functions so permissions and lineage come with them. |
| RetrievalTurning a question into passages the caller is allowed to see. | Open source pgvector in the same PostgreSQLdecided, not built | Cloud service Vector Search; embeddings in Delta plus a function where Vector Search is refuseddecided, not built | Open source pgvector in the same RDS PostgreSQLdecided, not built | Open source pgvector on Cloud SQL or AlloyDB, or Vertex AI Vector Searchproposed, no stack file | Open source Not yet bound (pgvector is the natural fit)decided, not built | Start with pgvector in the database you already run: one store, one backup. Use Vector Search when the corpus already lives in Delta. | |
| State and long-running work | StateConversation, checkpoints, idempotency records, case records. | Cloud service Azure Database for PostgreSQLdecided, not built | Cloud service Delta tables in Unity Catalog (Lakebase when row updates must be fast)built and run | Cloud service RDS for PostgreSQLdecided, not built | Cloud service Cloud SQL or AlloyDB for PostgreSQLproposed, no stack file | Open source PostgreSQLbuilt and run | Managed PostgreSQL of your cloud: the same engine on every stack, so the code does not change. On Databricks, Delta. |
| WorkflowDurable steps and waits that survive a restart. | Open source DBOS, a library on the same PostgreSQLdecided, not built | Cloud service Lakeflow Jobsdecided, not built | Open source DBOS on the same RDSdecided, not built | Open source DBOS on Cloud SQL (Cloud Workflows is the managed alternative)proposed, no stack file | Open source Temporalbuilt and run | DBOS by default: no new infrastructure. Temporal when undo flows grow. Lakeflow Jobs on Databricks, where a wait is a run that ends and a new run that starts on the decision. | |
| ApprovalWhere an irreversible action waits for a person. | Open source DBOS workflowsdecided, not built | Cloud service A Databricks App writing the decision to a Delta tabledecided, not built | Open source DBOS workflowsdecided, not built | Open source DBOS workflowsproposed, no stack file | Open source Temporal updatesbuilt and run | Follow the workflow choice. The decision is a message that your own rule validates, never a button the model can press. | |
| TriggerWhat starts a run nobody is watching, once. | Built by us Requests storedecided, not built | Built by us Requests storedecided, not built | Built by us Requests storedecided, not built | Built by us Requests storeproposed, no stack file | Built by us Requests storebuilt and run | Build it. Exactly-once starts are a rule about your request ids. | |
| ChannelThe chat surface and the desk where a person takes over. | Open source Chatwoot agent botdecided, not built | Open source Chatwoot agent botdecided, not built | Open source Chatwoot agent botdecided, not built | Open source Chatwoot agent botproposed, no stack file | Open source Chatwoot agent botbuilt and run | Open source, same everywhere. Skip it if the agent is a job with no chat. | |
| Trust and policy | IdentityWho the unit of work acts for, carried to every tool. | Cloud service Entra ID (External ID for customers)decided, not built | Cloud service Unity Catalog service principal; reviewer sign-in from the workspacedecided, not built | Cloud service Cognitodecided, not built | Cloud service Cloud IAM with Identity Platformproposed, no stack file | Open source Keycloakbuilt and run | Your cloud's identity service. Keycloak only where there is no cloud identity to lean on. |
| SecretsA credential reference that becomes a secret only at the moment of use. | Cloud service Key Vaultdecided, not built | Cloud service Secret scopesdecided, not built | Cloud service Secrets Managerdecided, not built | Cloud service Secret Managerproposed, no stack file | Built by us Environment settingsbuilt and run | Your cloud's secret store. Environment settings are acceptable for a laptop only. | |
| PolicyRules checked at declared points: before the model, after it, before a tool. | Built by us Positioned rule listsdecided, not built | Built by us Positioned rule listsdecided, not built | Built by us Positioned rule listsdecided, not built | Built by us Positioned rule listsproposed, no stack file | Built by us Positioned rule listsbuilt and run | Build it. No guardrail product knows that a draft's number must equal a tool's number. Cloud filters sit in front as a second layer. | |
| Observe, record and prove | TelemetryOne span per call and one identifier for the whole task. | Cloud service Application Insights through OpenTelemetrydecided, not built | Cloud service MLflow tracingdecided, not built | Cloud service OpenTelemetry to CloudWatchdecided, not built | Cloud service OpenTelemetry to Cloud Trace and Loggingproposed, no stack file | Open source OpenTelemetry to Langfusebuilt and run | Emit OpenTelemetry everywhere; take the cloud's backend. Langfuse is a good choice off-cloud. |
| RecorderCapture a real interaction and replay it without a network. | Built by us JSON cassettesdecided, not built | Built by us JSON cassettesdecided, not built | Built by us JSON cassettesdecided, not built | Built by us JSON cassettesproposed, no stack file | Built by us JSON cassettes (VCR library is the target)built and run | Build it small; swap to a recording library when it earns its place. | |
| Evaluation entry pointA callable the evaluations drive, with graders injected. | Open source AgentTwindecided, not built | Open source AgentTwinbuilt and run | Open source AgentTwindecided, not built | Open source AgentTwinproposed, no stack file | Open source AgentTwinbuilt and run | AgentTwin on every stack: scenarios are written once, against a declared world. | |
| Release and configuration | ConfigurationModel pins, flags, thresholds, allow-lists, outside the image. | Cloud service App Configurationdecided, not built | Cloud service Bundle variables and job parametersdecided, not built | Cloud service AppConfigdecided, not built | Cloud service A managed config service, or settings in Gitproposed, no stack file | Open source Typed settings (OpenFeature is the target)built and run | Your cloud's config service. Keep the numbers in the specification and read them, never restate them. |
| Agent registryThe catalogue of agents and tools that may run. | Cloud service API Centerdecided, not built | Cloud service Unity Catalog models and functionsdecided, not built | Cloud service Agent Registry (not offered in every region; a catalogue in Git works)decided, not built | Cloud service Artifact Registry plus a catalogue in Gitproposed, no stack file | Open source A catalogue in Gitbuilt and run | A catalogue in Git is enough to start. Use the platform registry when it gives you permissions you would otherwise build. |
The same map, cut three ways
The first table says which product fills each concern. This one asks a different question for each platform: for every concern, how much does the platform hand you, how much comes from assets we already share across all our agents, and what is left that every single agent must write for itself.
Boxes show who carries the concern: platform, shared assets, agent. A darker first box means the platform carries more.
Azure
| Concern | Who carries it | X · The platform gives out of the box | Y · Shared assets we reuse | Z · Each agent writes |
|---|---|---|---|---|
| Control loop | platform nothing for this | Nothing: it hosts the container | The loop: step and cost budgets, repeat detection, typed endingsshared asset is the whole answer | Its budget numbers from the specification, its standing instruction, and what each ending does next |
| Hosting | platform carries most of it | Container Apps with revisions for canary releases | A service entry point and settings loaderadapter ready: container build | Its container or job definition and its environment settings |
| Model access | platform carries most of it | API Management as gateway; Azure's model service | A model port with an OpenAI-compatible client, retries, a breaker and an approved-model listadapter ready: OpenAI-compatible client | Which models are approved, their prices, and sampling settings |
| Tools | platform nothing for this | Nothing specific; you host MCP servers | A tool client over MCP: result size limits, side-effect classes, idempotency keys, identity on every calladapter ready: MCP client | Its own tools: names, input and output shapes, which are read or irreversible, who may call each |
| Knowledge retrieval | platform carries part of it | AI Search exists; not chosen, pgvector is | A retrieval port: the caller's rights applied before ranking, a source on every passageadapter ready: pgvector | Its corpus, how it is cut into passages, how sources are linked |
| State and records | platform carries most of it | Azure Database for PostgreSQL | State stores: checkpoints, idempotency records, facts, conversationadapter ready: PostgreSQL store | Its own entities and their migrations (for example the drafts and decisions it keeps) |
| Long waits and approvals | platform carries part of it | Durable Functions and Logic Apps exist; not chosen | Approval and escalation machinery: durable waits, expiry, a decision validated by a rule, a queue for a personadapter ready: DBOS workflows | Who may decide, how long a decision stays valid, and which conditions escalate (all from its specification) |
| Identity and access | platform carries most of it | Entra ID; managed identity | An identity carried into every tool call; the far end verifies itadapter ready: token exchange | Its scopes, and which rows each caller may see or change |
| Secrets | platform carries most of it | Key Vault | A secrets port: a reference becomes a secret only at the moment of useadapter still to write: Key Vault adapter | Nothing but the names of the secrets it needs |
| Policy and guardrails | platform carries part of it | Content Safety filters: generic | Rule positions (before the model, after it, before a tool), output rules, redactionshared asset is the whole answer | Its own rules: what it refuses, which checks a draft must pass (for example a number must equal a tool's number) |
| Telemetry | platform carries most of it | Application Insights | One span per call, one identifier per task, redaction before export, standard span namesadapter ready: Azure Monitor distro | Which domain facts it records, and the labels its dashboards need |
| Evaluation | platform carries part of it | Foundry evaluations: model-level | AgentTwin (worlds, scenarios, simulated users), the evaluation library, the four gates, the cataloguesshared asset is the whole answer | Its world, its scenarios, its labelled cases and its domain checks |
| Configuration and release | platform carries most of it | App Configuration | A profile loader: decisions, thresholds and accepted gaps resolved from a stack fileadapter still to write: config adapter | Its profile answers and its own thresholds; it never restates the specification's numbers |
| Cost | platform carries part of it | Cost Management: by resource, not by task | A per-task meter, a price table, a ceiling that stops workshared asset is the whole answer | Its ceiling per task and what counts as a successful task |
Databricks
| Concern | Who carries it | X · The platform gives out of the box | Y · Shared assets we reuse | Z · Each agent writes |
|---|---|---|---|---|
| Control loop | platform carries part of it | Agent Framework interface for logging and serving; the loop stays ours | The loop: step and cost budgets, repeat detection, typed endingsshared asset is the whole answer | Its budget numbers from the specification, its standing instruction, and what each ending does next |
| Hosting | platform carries most of it | Lakeflow jobs; Databricks Apps for a desk | A service entry point and settings loaderadapter still to write: job wrapper | Its container or job definition and its environment settings |
| Model access | platform carries most of it | Foundation Model APIs; AI Gateway where a serving endpoint is allowed | A model port with an OpenAI-compatible client, retries, a breaker and an approved-model listadapter ready: OpenAI-compatible client (used) | Which models are approved, their prices, and sampling settings |
| Tools | platform carries most of it | Unity Catalog functions: typed, permissioned, audited | A tool client over MCP: result size limits, side-effect classes, idempotency keys, identity on every calladapter still to write: tool client over UC functions | Its own tools: names, input and output shapes, which are read or irreversible, who may call each |
| Knowledge retrieval | platform carries most of it | Vector Search, synced from Delta | A retrieval port: the caller's rights applied before ranking, a source on every passageadapter still to write: Vector Search adapter | Its corpus, how it is cut into passages, how sources are linked |
| State and records | platform carries most of it | Delta tables in Unity Catalog | State stores: checkpoints, idempotency records, facts, conversationadapter still to write: Delta store | Its own entities and their migrations (for example the drafts and decisions it keeps) |
| Long waits and approvals | platform carries part of it | Lakeflow Jobs run steps; a wait is a job that ends and restarts | Approval and escalation machinery: durable waits, expiry, a decision validated by a rule, a queue for a personadapter still to write: job per decision | Who may decide, how long a decision stays valid, and which conditions escalate (all from its specification) |
| Identity and access | platform carries most of it | Unity Catalog service principal and workspace sign-in | An identity carried into every tool call; the far end verifies itadapter still to write: workspace identity | Its scopes, and which rows each caller may see or change |
| Secrets | platform carries most of it | Secret scopes | A secrets port: a reference becomes a secret only at the moment of useadapter still to write: secret scope adapter | Nothing but the names of the secrets it needs |
| Policy and guardrails | platform carries part of it | AI Gateway safety and PII options: generic | Rule positions (before the model, after it, before a tool), output rules, redactionshared asset is the whole answer | Its own rules: what it refuses, which checks a draft must pass (for example a number must equal a tool's number) |
| Telemetry | platform carries most of it | MLflow tracing | One span per call, one identifier per task, redaction before export, standard span namesadapter still to write: MLflow exporter | Which domain facts it records, and the labels its dashboards need |
| Evaluation | platform carries part of it | Agent Evaluation and MLflow evaluation: model-level | AgentTwin (worlds, scenarios, simulated users), the evaluation library, the four gates, the cataloguesshared asset is the whole answer | Its world, its scenarios, its labelled cases and its domain checks |
| Configuration and release | platform carries most of it | Bundle variables and job parameters | A profile loader: decisions, thresholds and accepted gaps resolved from a stack fileadapter still to write: bundle adapter | Its profile answers and its own thresholds; it never restates the specification's numbers |
| Cost | platform carries part of it | Billing system tables: by resource, not by task | A per-task meter, a price table, a ceiling that stops workshared asset is the whole answer | Its ceiling per task and what counts as a successful task |
AWS
| Concern | Who carries it | X · The platform gives out of the box | Y · Shared assets we reuse | Z · Each agent writes |
|---|---|---|---|---|
| Control loop | platform carries part of it | AgentCore Runtime and Bedrock Agents exist; deliberately not used for the loop | The loop: step and cost budgets, repeat detection, typed endingsshared asset is the whole answer | Its budget numbers from the specification, its standing instruction, and what each ending does next |
| Hosting | platform carries most of it | ECS on Fargate | A service entry point and settings loaderadapter ready: container build | Its container or job definition and its environment settings |
| Model access | platform carries most of it | Bedrock; AgentCore Gateway | A model port with an OpenAI-compatible client, retries, a breaker and an approved-model listadapter ready: OpenAI-compatible client | Which models are approved, their prices, and sampling settings |
| Tools | platform carries part of it | AgentCore Gateway turns APIs into tools | A tool client over MCP: result size limits, side-effect classes, idempotency keys, identity on every calladapter ready: MCP client | Its own tools: names, input and output shapes, which are read or irreversible, who may call each |
| Knowledge retrieval | platform carries part of it | Knowledge Bases exist; not chosen, pgvector is | A retrieval port: the caller's rights applied before ranking, a source on every passageadapter ready: pgvector | Its corpus, how it is cut into passages, how sources are linked |
| State and records | platform carries most of it | RDS for PostgreSQL | State stores: checkpoints, idempotency records, facts, conversationadapter ready: PostgreSQL store | Its own entities and their migrations (for example the drafts and decisions it keeps) |
| Long waits and approvals | platform carries part of it | Step Functions exist; not chosen | Approval and escalation machinery: durable waits, expiry, a decision validated by a rule, a queue for a personadapter ready: DBOS workflows | Who may decide, how long a decision stays valid, and which conditions escalate (all from its specification) |
| Identity and access | platform carries most of it | Cognito and IAM | An identity carried into every tool call; the far end verifies itadapter ready: signing helpers | Its scopes, and which rows each caller may see or change |
| Secrets | platform carries most of it | Secrets Manager and Parameter Store | A secrets port: a reference becomes a secret only at the moment of useadapter ready: AWS adapter | Nothing but the names of the secrets it needs |
| Policy and guardrails | platform carries part of it | Bedrock Guardrails: generic | Rule positions (before the model, after it, before a tool), output rules, redactionshared asset is the whole answer | Its own rules: what it refuses, which checks a draft must pass (for example a number must equal a tool's number) |
| Telemetry | platform carries most of it | CloudWatch; X-Ray | One span per call, one identifier per task, redaction before export, standard span namesadapter ready: OpenTelemetry export | Which domain facts it records, and the labels its dashboards need |
| Evaluation | platform carries part of it | Bedrock evaluations: model-level | AgentTwin (worlds, scenarios, simulated users), the evaluation library, the four gates, the cataloguesshared asset is the whole answer | Its world, its scenarios, its labelled cases and its domain checks |
| Configuration and release | platform carries most of it | AppConfig | A profile loader: decisions, thresholds and accepted gaps resolved from a stack fileadapter ready: AWS adapter | Its profile answers and its own thresholds; it never restates the specification's numbers |
| Cost | platform carries part of it | Cost Explorer: by resource, not by task | A per-task meter, a price table, a ceiling that stops workshared asset is the whole answer | Its ceiling per task and what counts as a successful task |
GCP (proposed)
| Concern | Who carries it | X · The platform gives out of the box | Y · Shared assets we reuse | Z · Each agent writes |
|---|---|---|---|---|
| Control loop | platform carries part of it | Agent runtimes exist (ADK, Agent Engine); would not own the loop | The loop: step and cost budgets, repeat detection, typed endingsshared asset is the whole answer | Its budget numbers from the specification, its standing instruction, and what each ending does next |
| Hosting | platform carries most of it | Cloud Run | A service entry point and settings loaderadapter ready: container build | Its container or job definition and its environment settings |
| Model access | platform carries most of it | Vertex AI model catalogue; Apigee as gateway | A model port with an OpenAI-compatible client, retries, a breaker and an approved-model listadapter still to write: check Vertex's compatible endpoint | Which models are approved, their prices, and sampling settings |
| Tools | platform nothing for this | Nothing specific; you host MCP servers | A tool client over MCP: result size limits, side-effect classes, idempotency keys, identity on every calladapter ready: MCP client | Its own tools: names, input and output shapes, which are read or irreversible, who may call each |
| Knowledge retrieval | platform carries part of it | Vertex AI Vector Search and Search exist | A retrieval port: the caller's rights applied before ranking, a source on every passageadapter ready: pgvector on Cloud SQL | Its corpus, how it is cut into passages, how sources are linked |
| State and records | platform carries most of it | Cloud SQL or AlloyDB for PostgreSQL | State stores: checkpoints, idempotency records, facts, conversationadapter ready: PostgreSQL store | Its own entities and their migrations (for example the drafts and decisions it keeps) |
| Long waits and approvals | platform carries part of it | Cloud Workflows exists | Approval and escalation machinery: durable waits, expiry, a decision validated by a rule, a queue for a personadapter ready: DBOS on Cloud SQL | Who may decide, how long a decision stays valid, and which conditions escalate (all from its specification) |
| Identity and access | platform carries most of it | Cloud IAM; Identity Platform | An identity carried into every tool call; the far end verifies itadapter still to write: Google identity | Its scopes, and which rows each caller may see or change |
| Secrets | platform carries most of it | Secret Manager | A secrets port: a reference becomes a secret only at the moment of useadapter still to write: Secret Manager adapter | Nothing but the names of the secrets it needs |
| Policy and guardrails | platform carries part of it | Vertex safety filters and Model Armor: generic | Rule positions (before the model, after it, before a tool), output rules, redactionshared asset is the whole answer | Its own rules: what it refuses, which checks a draft must pass (for example a number must equal a tool's number) |
| Telemetry | platform carries most of it | Cloud Trace and Logging | One span per call, one identifier per task, redaction before export, standard span namesadapter ready: OpenTelemetry export | Which domain facts it records, and the labels its dashboards need |
| Evaluation | platform carries part of it | Gen AI evaluation service: model-level | AgentTwin (worlds, scenarios, simulated users), the evaluation library, the four gates, the cataloguesshared asset is the whole answer | Its world, its scenarios, its labelled cases and its domain checks |
| Configuration and release | platform carries part of it | A managed config service or Git | A profile loader: decisions, thresholds and accepted gaps resolved from a stack fileadapter still to write: config adapter | Its profile answers and its own thresholds; it never restates the specification's numbers |
| Cost | platform carries part of it | Billing export: by resource, not by task | A per-task meter, a price table, a ceiling that stops workshared asset is the whole answer | Its ceiling per task and what counts as a successful task |
Open stack
| Concern | Who carries it | X · The platform gives out of the box | Y · Shared assets we reuse | Z · Each agent writes |
|---|---|---|---|---|
| Control loop | platform nothing for this | Nothing | The loop: step and cost budgets, repeat detection, typed endingsshared asset is the whole answer | Its budget numbers from the specification, its standing instruction, and what each ending does next |
| Hosting | platform carries part of it | You run the containers | A service entry point and settings loaderadapter ready: compose file | Its container or job definition and its environment settings |
| Model access | platform nothing for this | Nothing; LiteLLM is the open-source gateway you run | A model port with an OpenAI-compatible client, retries, a breaker and an approved-model listadapter ready: OpenAI-compatible client | Which models are approved, their prices, and sampling settings |
| Tools | platform nothing for this | Nothing | A tool client over MCP: result size limits, side-effect classes, idempotency keys, identity on every calladapter ready: MCP client | Its own tools: names, input and output shapes, which are read or irreversible, who may call each |
| Knowledge retrieval | platform nothing for this | Nothing | A retrieval port: the caller's rights applied before ranking, a source on every passageadapter ready: pgvector | Its corpus, how it is cut into passages, how sources are linked |
| State and records | platform carries part of it | You run PostgreSQL | State stores: checkpoints, idempotency records, facts, conversationadapter ready: PostgreSQL store | Its own entities and their migrations (for example the drafts and decisions it keeps) |
| Long waits and approvals | platform nothing for this | Nothing; Temporal is the open-source choice | Approval and escalation machinery: durable waits, expiry, a decision validated by a rule, a queue for a personadapter ready: Temporal | Who may decide, how long a decision stays valid, and which conditions escalate (all from its specification) |
| Identity and access | platform nothing for this | Nothing; Keycloak is the open-source choice | An identity carried into every tool call; the far end verifies itadapter ready: token check | Its scopes, and which rows each caller may see or change |
| Secrets | platform nothing for this | Nothing | A secrets port: a reference becomes a secret only at the moment of useadapter ready: environment settings | Nothing but the names of the secrets it needs |
| Policy and guardrails | platform nothing for this | Nothing | Rule positions (before the model, after it, before a tool), output rules, redactionshared asset is the whole answer | Its own rules: what it refuses, which checks a draft must pass (for example a number must equal a tool's number) |
| Telemetry | platform carries part of it | You run Langfuse or a collector | One span per call, one identifier per task, redaction before export, standard span namesadapter ready: OpenTelemetry export | Which domain facts it records, and the labels its dashboards need |
| Evaluation | platform nothing for this | Nothing | AgentTwin (worlds, scenarios, simulated users), the evaluation library, the four gates, the cataloguesshared asset is the whole answer | Its world, its scenarios, its labelled cases and its domain checks |
| Configuration and release | platform nothing for this | Nothing | A profile loader: decisions, thresholds and accepted gaps resolved from a stack fileadapter ready: typed settings | Its profile answers and its own thresholds; it never restates the specification's numbers |
| Cost | platform nothing for this | Nothing | A per-task meter, a price table, a ceiling that stops workshared asset is the whole answer | Its ceiling per task and what counts as a successful task |
The count, per platform
| Platform | Concerns the platform mostly carries (of 14) | Concerns it gives nothing for | Adapters ready in the shared library | Adapters still to write |
|---|---|---|---|---|
| Azure | 7 | 2 | 8 | 2 |
| Databricks | 9 | 0 | 1 | 9 |
| AWS | 7 | 0 | 10 | 0 |
| GCP (proposed) | 6 | 1 | 6 | 4 |
| Open stack | 0 | 11 | 10 | 0 |
The levels in this view are our reading of each platform, not a measurement. The adapter status comes from the shared library as of 20 November 2026; GCP and Databricks adapters do not exist yet. Platform product lists change quickly.
How to decide, in three rules
What to check before relying on this
- The GCP column. It is our proposal, not a tested stack. Confirm each service against Google's current documentation and region availability.
- Azure and AWS cells. They record decisions. Few have been built against the real service.
- Databricks limits. On the free edition we found Vector Search and custom model serving refused, which also rules out AI Gateway on a custom endpoint. A paid workspace may differ.
- Prices and tiers. Not shown here on purpose. Get them from each vendor for your volume.
Where to go next
- Agent Harness: the course, concern by concern.
- The agent harness on Databricks and on AWS.
- AI gateways: which one to use, the model-access row in depth.
- Open spec: AI Harness Catalog, where the ports are defined.