Official OpenAI Agents backend, current AI SDK frontend

This slice proves the narrowest official integration path: OpenAI Agents SDK on the server, the official AI SDK UI bridge in the route, and the existing chat workspace left intact.

Readyopenai/gpt-5-mini

OpenAI Agents workspace is ready

Send a prompt and exercise the official OpenAI Agents SDK to AI SDK UI bridge without leaving the current frontend stack.

Voice

Browser lane.

Disconnected

setup-required

Runtime

v24.18.0

Contract

This slot keeps the agent backend on the official OpenAI run path and lets the existing AI SDK UI frontend consume the result without a custom stream protocol.

Running

Workflowopenai-agents-sdk-demo

Continuationprevious-response-id-or-memory-session

Max turns8

Abortrequest.signal

Last responseNot available yet

Context

PrimitiveRunContext

Session bindingsession-id

Run modeNo run yet

Default targetTesla

Channels

dynamic instructionstool()guardrailsagent.asTool toolInput

Latest user prompt

No context metadata yet

Sessions

PrimitiveMemorySession

Storageprocess-local

Transportassistant-message metadata

Session idNot created yet

History items0

CRUD helpersget/add/pop/clear

Streaming

BridgecreateAiSdkUiMessageStream()

Raw model events0

Run item events0

Agents

No stream metadata yet

Raw event types

No stream metadata yet

Run item names

No stream metadata yet

Sources

No stream metadata yet

AI SDK Extension

UI bridgecreateAiSdkUiMessageStream()

Response helpercreateAiSdkUiMessageStreamResponse()

Bridge statusconfigured

Model adapteraisdk(model)

Adapter statusnot-used

Official docs recommend the default OpenAI provider for OpenAI models, and the beta aisdk() model adapter does not support deferred Responses tool loading.

Voice Agents

RealtimeAgentRealtimeSessionWebRTCsetup-required

Browser modelgpt-realtime-2

Voicemarin

Tools2 total / 1 gated

Handoffs1

Provider lanes2

Workspacesupported

Handoffs

Configured2

Active agentNo handoff yet

Configured handoffs

Market Context AgentResearch Lead handoff

Targets

No handoff metadata yet

Transitions

No handoff metadata yet

MCP

Transportstreamable-http

LifecycleconnectMcpServers

Route/api/demos/openai-agents-sdk-demo/mcp

Strict schemasenabled

Active servers0

Failed servers0

SDK primitives

MCPServerStreamableHttpconnectMcpServersAgent.mcpServersAgent.mcpConfig

Configured tools

openai_agents_demo_docs__read_demo_docopenai_agents_demo_docs__search_demo_docs

Used this run

No MCP tool usage yet

Failed server errors

No MCP connection errors

Sandbox

Clientunix_local

Agent modelopenai/gpt-5.4-mini

Manifest root/workspace

WorkspacelocalDir() -> temp workspace

Persistencesession-id -> process-local sessionState

Active agentNo sandbox run yet

Workspace readyunknown

Mounted paths

/workspace/docs/workspace/feature

Default capabilities

filesystem()shell()compaction()

SDK primitives

SandboxAgentManifestCapabilities.default()UnixLocalSandboxClientRunConfig.sandbox

Persisted sessions

Tracing

Runtime defaultenabled

Workflow sourceRunConfig.workflowName

Group strategysession-id

Tracingenabled

Sensitive dataincluded

Export keymissing

Trace idNo trace yet

Group idNo group yet

SDK primitives

generateTraceIdrun({ workflowName, traceId, groupId, traceMetadata, tracingDisabled, traceIncludeSensitiveData, tracing })

Trace metadata keys

No trace metadata yet

Disable switch

OPENAI_AGENTS_DISABLE_TRACING

Model

Modelopenai/gpt-5-mini

APIresponses

Transporthttp

Reasoningmedium

Verbositylow

Providergateway-openai-client

Guardrails

prompt_scope_guardrail

defineInputGuardrail()

configured

Blocks direct prompt-injection or system-prompt extraction requests before the run proceeds.

investment_advice_guardrail

defineOutputGuardrail()

configured

Blocks direct buy or sell recommendation phrasing in the final agent output.

Approvals

Pending0

Decisions0

Paused statenone

Pending approvals

No pending approval metadata yet

Decisions

No approval decisions yet

Results

Active agentNo settled result yet

Last agentNo settled result yet

History0

Output items0

New items0

Raw responses0

Interruptions0

Requests0

Tokens0

Run statenot captured

Final output

No settled result preview yet

Tools

build_research_brief

tool()

configured

Thin official tool() example for the Tesla-style research flow. The current AI Gateway path can reject the SDK's default run_llm_again continuation after local function tools; the demo surfaces that provider error instead of injecting a fake final answer.

publish_research_summary

tool({ needsApproval: true })

configured

Official tool({ needsApproval: true }) path for a human approval interruption before sharing research externally. The current AI Gateway path can reject the SDK's default run_llm_again continuation after local function tools; the demo surfaces that provider error instead of injecting a fake final answer.

draft_financial_follow_up

tool({ deferLoading: true })

setup-required

Requires an OpenAI Responses model with tool_search support, such as openai/gpt-5.4-mini or newer.

build_risk_watchlist

tool({ deferLoading: true })

setup-required

Requires an OpenAI Responses model with tool_search support, such as openai/gpt-5.4-mini or newer.

web_search

webSearchTool()

configured

Hosted OpenAI search through the current AI Gateway OpenAI-compatible Responses path.

code_interpreter

codeInterpreterTool()

configured

Hosted OpenAI code interpreter for math, tables, and quick analysis.

research_memo_agent

agent.asTool()

configured

Specialist sub-agent exposed through agent.asTool() for memo synthesis.

sandbox_workspace_agent

SandboxAgent.asTool()

configured

Official SandboxAgent specialist mounted over the demo docs and feature slice, then exposed through agent.asTool().

file_search

fileSearchTool()

setup-required

Set OPENAI_AGENTS_VECTOR_STORE_IDS to one or more comma-separated OpenAI vector store ids.

image_generation

imageGenerationTool()

provider-blocked

Registered through imageGenerationTool(), but blocked on the current AI Gateway Responses path. The hosted image tool can complete without a renderable image artifact, and streamed runs can terminate before the AI SDK UI bridge receives a usable file part.

tool_search

toolSearchTool()

setup-required

Requires an OpenAI Responses model with tool_search support, such as openai/gpt-5.4-mini or newer.

Guide Coverage

Agents

Agent

Implemented

Agent instance passed to run()

https://openai.github.io/openai-agents-js/guides/agents/

Run: ReadyProvider: available

Models

model, modelSettings

Implemented

Agent modelSettings.reasoning + text.verbosity on the responses API path

https://openai.github.io/openai-agents-js/guides/models/

Run: ReadyProvider: available

Tools

tool()

Implemented

RunToolCallItem / RunToolCallOutputItem

https://openai.github.io/openai-agents-js/guides/tools/

Run: ReadyProvider: available

Guardrails

inputGuardrails / outputGuardrails

Implemented

Guardrail tripwire result

https://openai.github.io/openai-agents-js/guides/guardrails/

Run: ReadyProvider: available

Running Agents

run()

Implemented

run() + previousResponseId / MemorySession continuation + maxTurns + AbortSignal

https://openai.github.io/openai-agents-js/guides/running-agents/

Run: ReadyProvider: available

Streaming

RunStreamEvent

Implemented

RunStreamEvent metadata from raw_model_stream_event, run_item_stream_event, and agent_updated_stream_event

https://openai.github.io/openai-agents-js/guides/streaming/

Run: ReadyProvider: available

Agent Orchestration

agent.asTool()

Implemented

agent.asTool() invocation through a specialist sub-agent

https://openai.github.io/openai-agents-js/guides/multi-agent/

Run: ReadyProvider: available

Handoffs

handoff()

Implemented

RunHandoffCallItem / RunHandoffOutputItem

https://openai.github.io/openai-agents-js/guides/handoffs/

Run: ReadyProvider: available

Results

RunResult

Implemented

finalOutput, history, newItems, state

https://openai.github.io/openai-agents-js/guides/results/

Run: ReadyProvider: available

Human-in-the-loop

interruptions / approval

Implemented

RunToolApprovalItem interruption

https://openai.github.io/openai-agents-js/guides/human-in-the-loop/

Run: ReadyProvider: available

Sessions

MemorySession

Implemented

MemorySession history + assistant metadata session id

https://openai.github.io/openai-agents-js/guides/sessions/

Run: ReadyProvider: available

Context Management

RunContext<T>

Implemented

RunContext through run(), dynamic instructions, tools, and guardrails

https://openai.github.io/openai-agents-js/guides/context/

Run: ReadyProvider: available

MCP

MCPServerStreamableHttp / connectMcpServers

Implemented

MCP server connection state + server-prefixed MCP tool call items

https://openai.github.io/openai-agents-js/guides/mcp/

Run: ReadyProvider: available

Tracing

traceId / groupId / RunConfig.tracing

Implemented

workflowName + traceId + groupId + traceMetadata + tracingDisabled

https://openai.github.io/openai-agents-js/guides/tracing/

Run: ReadyProvider: available

Sandbox Agents

SandboxAgent / RunConfig.sandbox

Implemented

SandboxAgent lifecycle + RunConfig.sandbox + persisted sandbox session state

https://openai.github.io/openai-agents-js/guides/sandbox-agents/

Run: ReadyProvider: available

Voice Agents

RealtimeAgent / RealtimeSession

Implemented

RealtimeSession.connect({ apiKey }) over WebRTC plus visible browser microphone controls

https://openai.github.io/openai-agents-js/guides/voice-agents/

Run: BlockedProvider: setup-required

AI SDK Extension

createAiSdkUiMessageStream() / aisdk(model)

Implemented

AI SDK UI stream from createAiSdkUiMessageStream(); aisdk(model) adapter boundary is explicit

https://openai.github.io/openai-agents-js/extensions/ai-sdk/

Run: ReadyProvider: available

Agent Demos

Opening demo

The demo workspace is loading. This can take a little longer on slow mobile networks.

Official OpenAI Agents backend, current AI SDK frontend

This slice proves the narrowest official integration path: OpenAI Agents SDK on the server, the official AI SDK UI bridge in the route, and the existing chat workspace left intact.

Readyopenai/gpt-5-mini

OpenAI Agents workspace is ready

Send a prompt and exercise the official OpenAI Agents SDK to AI SDK UI bridge without leaving the current frontend stack.

Voice

Browser lane.

Disconnected

setup-required

Runtime

v24.18.0

Contract

This slot keeps the agent backend on the official OpenAI run path and lets the existing AI SDK UI frontend consume the result without a custom stream protocol.

Running

Workflowopenai-agents-sdk-demo

Continuationprevious-response-id-or-memory-session

Max turns8

Abortrequest.signal

Last responseNot available yet

Context

PrimitiveRunContext

Session bindingsession-id

Run modeNo run yet

Default targetTesla

Channels

dynamic instructionstool()guardrailsagent.asTool toolInput

Latest user prompt

No context metadata yet

Sessions

PrimitiveMemorySession

Storageprocess-local

Transportassistant-message metadata

Session idNot created yet

History items0

CRUD helpersget/add/pop/clear

Streaming

BridgecreateAiSdkUiMessageStream()

Raw model events0

Run item events0

Agents

No stream metadata yet

Raw event types

No stream metadata yet

Run item names

No stream metadata yet

Sources

No stream metadata yet

AI SDK Extension

UI bridgecreateAiSdkUiMessageStream()

Response helpercreateAiSdkUiMessageStreamResponse()

Bridge statusconfigured

Model adapteraisdk(model)

Adapter statusnot-used

Official docs recommend the default OpenAI provider for OpenAI models, and the beta aisdk() model adapter does not support deferred Responses tool loading.

Voice Agents

RealtimeAgentRealtimeSessionWebRTCsetup-required

Browser modelgpt-realtime-2

Voicemarin

Tools2 total / 1 gated

Handoffs1

Provider lanes2

Workspacesupported

Handoffs

Configured2

Active agentNo handoff yet

Configured handoffs

Market Context AgentResearch Lead handoff

Targets

No handoff metadata yet

Transitions

No handoff metadata yet

MCP

Transportstreamable-http

LifecycleconnectMcpServers

Route/api/demos/openai-agents-sdk-demo/mcp

Strict schemasenabled

Active servers0

Failed servers0

SDK primitives

MCPServerStreamableHttpconnectMcpServersAgent.mcpServersAgent.mcpConfig

Configured tools

openai_agents_demo_docs__read_demo_docopenai_agents_demo_docs__search_demo_docs

Used this run

No MCP tool usage yet

Failed server errors

No MCP connection errors

Sandbox

Clientunix_local

Agent modelopenai/gpt-5.4-mini

Manifest root/workspace

WorkspacelocalDir() -> temp workspace

Persistencesession-id -> process-local sessionState

Active agentNo sandbox run yet

Workspace readyunknown

Mounted paths

/workspace/docs/workspace/feature

Default capabilities

filesystem()shell()compaction()

SDK primitives

SandboxAgentManifestCapabilities.default()UnixLocalSandboxClientRunConfig.sandbox

Persisted sessions

Tracing

Runtime defaultenabled

Workflow sourceRunConfig.workflowName

Group strategysession-id

Tracingenabled

Sensitive dataincluded

Export keymissing

Trace idNo trace yet

Group idNo group yet

SDK primitives

generateTraceIdrun({ workflowName, traceId, groupId, traceMetadata, tracingDisabled, traceIncludeSensitiveData, tracing })

Trace metadata keys

No trace metadata yet

Disable switch

OPENAI_AGENTS_DISABLE_TRACING

Model

Modelopenai/gpt-5-mini

APIresponses

Transporthttp

Reasoningmedium

Verbositylow

Providergateway-openai-client

Guardrails

prompt_scope_guardrail

defineInputGuardrail()

configured

Blocks direct prompt-injection or system-prompt extraction requests before the run proceeds.

investment_advice_guardrail

defineOutputGuardrail()

configured

Blocks direct buy or sell recommendation phrasing in the final agent output.

Approvals

Pending0

Decisions0

Paused statenone

Pending approvals

No pending approval metadata yet

Decisions

No approval decisions yet

Results

Active agentNo settled result yet

Last agentNo settled result yet

History0

Output items0

New items0

Raw responses0

Interruptions0

Requests0

Tokens0

Run statenot captured

Final output

No settled result preview yet

Tools

build_research_brief

tool()

configured

publish_research_summary

tool({ needsApproval: true })

configured

draft_financial_follow_up

tool({ deferLoading: true })

setup-required

Requires an OpenAI Responses model with tool_search support, such as openai/gpt-5.4-mini or newer.

build_risk_watchlist

tool({ deferLoading: true })

setup-required

Requires an OpenAI Responses model with tool_search support, such as openai/gpt-5.4-mini or newer.

web_search

webSearchTool()

configured

Hosted OpenAI search through the current AI Gateway OpenAI-compatible Responses path.

code_interpreter

codeInterpreterTool()

configured

Hosted OpenAI code interpreter for math, tables, and quick analysis.

research_memo_agent

agent.asTool()

configured

Specialist sub-agent exposed through agent.asTool() for memo synthesis.

sandbox_workspace_agent

SandboxAgent.asTool()

configured

Official SandboxAgent specialist mounted over the demo docs and feature slice, then exposed through agent.asTool().

file_search

fileSearchTool()

setup-required

Set OPENAI_AGENTS_VECTOR_STORE_IDS to one or more comma-separated OpenAI vector store ids.

image_generation

imageGenerationTool()

provider-blocked

tool_search

toolSearchTool()

setup-required

Requires an OpenAI Responses model with tool_search support, such as openai/gpt-5.4-mini or newer.

Guide Coverage

Agents

Agent

Implemented

Agent instance passed to run()

https://openai.github.io/openai-agents-js/guides/agents/

Run: ReadyProvider: available

Models

model, modelSettings

Implemented

Agent modelSettings.reasoning + text.verbosity on the responses API path

https://openai.github.io/openai-agents-js/guides/models/

Run: ReadyProvider: available

Tools

tool()

Implemented

RunToolCallItem / RunToolCallOutputItem

https://openai.github.io/openai-agents-js/guides/tools/

Run: ReadyProvider: available

Guardrails

inputGuardrails / outputGuardrails

Implemented

Guardrail tripwire result

https://openai.github.io/openai-agents-js/guides/guardrails/

Run: ReadyProvider: available

Running Agents

run()

Implemented

run() + previousResponseId / MemorySession continuation + maxTurns + AbortSignal

https://openai.github.io/openai-agents-js/guides/running-agents/

Run: ReadyProvider: available

Streaming

RunStreamEvent

Implemented

RunStreamEvent metadata from raw_model_stream_event, run_item_stream_event, and agent_updated_stream_event

https://openai.github.io/openai-agents-js/guides/streaming/

Run: ReadyProvider: available

Agent Orchestration

agent.asTool()

Implemented

agent.asTool() invocation through a specialist sub-agent

https://openai.github.io/openai-agents-js/guides/multi-agent/

Run: ReadyProvider: available

Handoffs

handoff()

Implemented

RunHandoffCallItem / RunHandoffOutputItem

https://openai.github.io/openai-agents-js/guides/handoffs/

Run: ReadyProvider: available

Results

RunResult

Implemented

finalOutput, history, newItems, state

https://openai.github.io/openai-agents-js/guides/results/

Run: ReadyProvider: available

Human-in-the-loop

interruptions / approval

Implemented

RunToolApprovalItem interruption

https://openai.github.io/openai-agents-js/guides/human-in-the-loop/

Run: ReadyProvider: available

Sessions

MemorySession

Implemented

MemorySession history + assistant metadata session id

https://openai.github.io/openai-agents-js/guides/sessions/

Run: ReadyProvider: available

Context Management

RunContext<T>

Implemented

RunContext through run(), dynamic instructions, tools, and guardrails

https://openai.github.io/openai-agents-js/guides/context/

Run: ReadyProvider: available

MCP

MCPServerStreamableHttp / connectMcpServers

Implemented

MCP server connection state + server-prefixed MCP tool call items

https://openai.github.io/openai-agents-js/guides/mcp/

Run: ReadyProvider: available

Tracing

traceId / groupId / RunConfig.tracing

Implemented

workflowName + traceId + groupId + traceMetadata + tracingDisabled

https://openai.github.io/openai-agents-js/guides/tracing/

Run: ReadyProvider: available

Sandbox Agents

SandboxAgent / RunConfig.sandbox

Implemented

SandboxAgent lifecycle + RunConfig.sandbox + persisted sandbox session state

https://openai.github.io/openai-agents-js/guides/sandbox-agents/

Run: ReadyProvider: available

Voice Agents

RealtimeAgent / RealtimeSession

Implemented

RealtimeSession.connect({ apiKey }) over WebRTC plus visible browser microphone controls

https://openai.github.io/openai-agents-js/guides/voice-agents/

Run: BlockedProvider: setup-required

AI SDK Extension

createAiSdkUiMessageStream() / aisdk(model)

Implemented

AI SDK UI stream from createAiSdkUiMessageStream(); aisdk(model) adapter boundary is explicit

https://openai.github.io/openai-agents-js/extensions/ai-sdk/

Run: ReadyProvider: available