API Reference

Braintrust experts can be invoked programmatically via REST API, enabling automated pipelines, CI/CD integrations, custom applications, and webhook-driven workflows.

Base URL

https://braintrust.ti.trilogy.com

Authentication

All API requests require an API key passed as the x-api-key header.

x-api-key: YOUR_API_KEY

Get your API key from Settings → Manage API Keys. See Managing API Keys for a step-by-step guide.

Never commit API keys to source control or log them in plain text. Treat them like passwords.

Invoke an Expert

POST /api/agents/invoke Send a message to an expert and receive a response. Supports streaming (default) and async job modes.

Request Body

Field	Type	Required	Description
`agent`	string	Yes*	Single agent ID (lowercase, hyphenated — from the expert’s page)
`agents`	string[]	Yes*	Array of agent IDs for multi-agent collaboration
`message`	string	Yes	Non-empty user message
`thread_token`	string	No	Conversation continuity handle (see Thread Continuity)
`stream`	boolean	No	Default `true`. Set `false` for async job mode
`context`	string	No	Extra background context injected into the conversation
`attachments`	Attachment[]	No	Array of file attachments uploaded via the presigned-URL flow. Each entry references a previously uploaded file by `s3_key` (returned by the upload endpoint). Each entry contains exactly `s3_key`, `type`, `name`. Maximum 20 files. See Attachments for the per-entry schema.
`deliver_to.gchat_space_id`	string	No	Deliver output to a Google Chat Space (e.g. `spaces/AAQAvU5RLdI`)
`deliver_to.agent_dm`	boolean	No	Deliver output as a DM from the agent’s identity

* Use either agent or agents, never both.

Fields async and wait_for_completion are not supported — use stream only.

Attachments

Each entry in the attachments array is an object with the following fields:

Field	Type	Required	Description
`s3_key`	string	Yes	Must start with `invoke-uploads/` and come from the presigned-URL flow below. Arbitrary S3 paths return 400.
`type`	string	Yes	MIME type of the uploaded file (e.g. `application/pdf`).
`name`	string	Yes	Original basename of the file.

No other fields are accepted — extra keys return 400. Maximum 20 attachments per invoke.

Files must be uploaded before invoking the expert. Use POST /api/agents/invoke/attachments/upload-url to obtain a presigned URL, upload the file to that URL, then include the returned attachment object (s3_key, type, name) in your invoke request. The s3_key must start with invoke-uploads/ — arbitrary S3 paths are rejected with 400.

JSON request body example with attachments:

{
  "agent": "data-analyst",
  "message": "Summarise the two reports attached.",
  "attachments": [
    {
      "s3_key": "invoke-uploads/you@example.com/session-1/1716300000000-q1-report.pdf",
      "type": "application/pdf",
      "name": "q1-report.pdf"
    },
    {
      "s3_key": "invoke-uploads/you@example.com/session-1/1716300000001-q2-report.pdf",
      "type": "application/pdf",
      "name": "q2-report.pdf"
    }
  ]
}

curl example:

curl -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": "data-analyst",
    "message": "Summarise the two reports attached.",
    "attachments": [
      {
        "s3_key": "invoke-uploads/you@example.com/session-1/1716300000000-q1-report.pdf",
        "type": "application/pdf",
        "name": "q1-report.pdf"
      },
      {
        "s3_key": "invoke-uploads/you@example.com/session-1/1716300000001-q2-report.pdf",
        "type": "application/pdf",
        "name": "q2-report.pdf"
      }
    ]
  }'

Get a presigned upload URL

Call this endpoint once per file before invoking with attachments. POST /api/agents/invoke/attachments/upload-url Request fields:

Field	Type	Required	Description
`uploadSessionId`	string	Yes	Stable client ID for the upload batch (e.g. UUID); scopes S3 keys per user. Note camelCase.
`filename`	string	Yes	Basename only — no `/` or `..`.
`contentType`	string	Yes	MIME type sent on the subsequent PUT (e.g. `application/pdf`).

Response fields:

Field	Type	Description
`uploadUrl`	string	Presigned S3 URL. PUT the file bytes to this URL with `Content-Type` matching `contentType`.
`s3Key`	string	The key to pass back as `s3_key` on invoke.
`expiresInSeconds`	number	URL lifetime in seconds (default 3600).
`attachment`	object	Ready-to-use attachment object containing `s3_key`, `type`, `name` — pass this directly in the invoke `attachments[]` array.

curl -s -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke/attachments/upload-url" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "uploadSessionId": "my-session-1",
    "filename": "report.pdf",
    "contentType": "application/pdf"
  }'

Response:

{
  "uploadUrl": "https://s3.amazonaws.com/braintrust-uploads/invoke-uploads/you@example.com/my-session-1/1716300000000-report.pdf?...",
  "s3Key": "invoke-uploads/you@example.com/my-session-1/1716300000000-report.pdf",
  "expiresInSeconds": 3600,
  "attachment": {
    "s3_key": "invoke-uploads/you@example.com/my-session-1/1716300000000-report.pdf",
    "type": "application/pdf",
    "name": "report.pdf"
  }
}

Reuse the same uploadSessionId across multiple files in one batch. Call upload-url once per file; each response gives a unique s3Key.

Streaming Mode (Default)

When stream is omitted or true, the response is a Server-Sent Events (SSE) stream at HTTP 200 with Content-Type: text/event-stream. Lines beginning with data: are JSON payloads. Lines beginning with : keepalive are heartbeat comments — ignore them when parsing. Streaming responses include an X-Request-Id response header containing the server-generated request correlation ID.

Log X-Request-Id from every streaming response alongside your task_id and thread_token. If a request behaves unexpectedly, the X-Request-Id value lets support engineers locate the exact server-side trace without needing full request payloads.

Event sequence

1. meta — always the first event

data: {"type":"meta","thread_token":"<uuid>","task_id":"<uuid>","request_id":"<uuid>","conversation_message_id":"<id>"}

2. chunk — zero or more token events

data: {"type":"chunk","actor":"Agent Display Name","text":"..."}

3. done — terminal success Single agent:

data: {"type":"done","status":"completed","output":"...","thread_token":"<uuid>","task_id":"<uuid>","request_id":"<uuid>","agent":"Agent Display Name"}

Multi-agent collaboration:

data: {"type":"done","status":"completed","output":"...","thread_token":"<uuid>","task_id":"<uuid>","request_id":"<uuid>","agents_used":["Agent A","Agent B"]}

On failure:

data: {"type":"error","error":"...","thread_token":"<uuid>","task_id":"<uuid>","request_id":"<uuid>"}

Every successful invoke turn gets a UUID request_id echoed in meta, done, async 202, poll responses, and the X-Request-Id response header. Save it to look up the LangSmith trace timeline (see Traces).

Example

curl -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent":"AGENT_ID","message":"What can you help me with?"}'

If you see meta → chunk lines → done, your integration is working. Save the thread_token from meta or done for follow-up messages.

Async Mode (`stream: false`)

When stream is false, the server returns 202 Accepted immediately with a task_id you can poll.

Step 1 — Create the async task

curl -s -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent":"AGENT_ID","message":"Summarize this project.","stream":false}'

Response:

{
  "status": "accepted",
  "task_id": "d7c615f6-...",
  "request_id": "aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee",
  "status_url": "/api/agents/invoke/d7c615f6-...",
  "thread_token": "a1b2c3d4-...",
  "conversation_message_id": "..."
}

Field	Meaning
`status`	Always `"accepted"` on 202
`task_id`	Use to poll `/api/agents/invoke/{task_id}`
`request_id`	UUID correlation ID for this turn — echoed in `meta`/`done` events, this 202 body, poll responses, and the `X-Request-Id` streaming response header. Use it with the trace timeline endpoint (see Traces).
`status_url`	Relative URL for polling; prepend the base URL
`thread_token`	Conversation continuity handle for follow-up messages
`conversation_message_id`	Internal message identifier

Step 2 — Poll for the result

curl -s "https://braintrust.ti.trilogy.com/api/agents/invoke/d7c615f6-..." \
  -H "x-api-key: YOUR_API_KEY"

Prepend the base URL to status_url to get the full polling URL. Possible status values: queued, running, completed, failed, timeout, canceled

Status	Response includes
`completed` or `timeout`	`output`, `agent` or `agents_used`, `duration_ms`
`failed`	`error`, `duration_ms`
`canceled`	includes `output` (empty string if no output was produced before cancel)

Step 3 — Cancel (optional)

Cancel any queued or running task (including a streaming run, using its task_id from meta):

curl -s -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke/{task_id}/cancel" \
  -H "x-api-key: YOUR_API_KEY"

Cancel is best-effort. On success you get 200:

{ "task_id": "...", "status": "canceled", "cancel_applied": true }

If the task is already in a terminal state, you still get 200 but with cancel_applied: false and a reason field:

reason	When it appears
`already_terminal`	Task already `completed`, `failed`, or `timeout`.
`already_canceled`	A prior cancel call already applied.
`already_completed_or_not_active`	Run finished or no longer active in the worker pool.

Other statuses:

Status	Meaning
`409`	Record exists but cannot be canceled (missing internal message id).
`404`	The `task_id` is not yours / not found.

Thread Continuity

thread_token keeps messages connected into a single conversation thread.

Scenario	What to do
First message, new thread	Omit `thread_token` — server mints a UUID and returns it in `meta` / `done` / `202`
Bring your own ID	Send any stable opaque string (ticket ID, session ID, etc.) — it is accepted as-is
Follow-up message	Send the same `thread_token` used on the first turn
Start over	Omit `thread_token` or use a new custom string

If you change the agents array while reusing the same thread_token, the conversation will not continue from the previous thread. Keep the same agent roster across follow-up messages to maintain continuity.

Multi-turn example

# First message — server mints the thread_token
curl -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent":"AGENT_ID","message":"Summarise last week sales data"}'

# Follow-up — continue the same thread
curl -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agent":"AGENT_ID","thread_token":"<token-from-first-response>","message":"Now break that down by region"}'

Multi-Agent Collaboration

Use the agents array (not agent) to orchestrate multiple experts in a single turn.

curl -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"agents":["archivist","prompt-to-skills"],"message":"Each introduce yourself."}'

The done event and poll responses include agents_used (display names of agents that contributed), not agents.
Keep the same agents roster and thread_token on follow-ups for continuity.

Delivery Options

Deliver to a Google Chat Space

curl -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent":"AGENT_ID",
    "message":"Post this update in the space.",
    "deliver_to":{"gchat_space_id":"spaces/SPACE_ID"}
  }'

Requirement: the BrainTrust bot must be a member of that space.

Deliver as DM from the agent

curl -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" \
  -H "x-api-key: YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent":"AGENT_ID",
    "message":"Send this as a DM.",
    "deliver_to":{"agent_dm":true}
  }'

Delivery diagnostics

Streaming done events may include delivery_error (string) when posting to the configured destination failed, or delivered_to (object) confirming where the result was sent. Polled GET task-status responses do NOT carry a separate delivery_error field on success — use streaming if you need delivery diagnostics in the HTTP body.

Read Endpoints

Task status

GET /api/agents/invoke/{task_id} Retrieve status and output for any invocation (streaming or async) by task_id. The task_id is returned in the streaming meta event and in the async 202 response.

curl -s "https://braintrust.ti.trilogy.com/api/agents/invoke/TASK_ID" \
  -H "x-api-key: YOUR_API_KEY"

Response example:

{
  "task_id": "d7c615f6-...",
  "request_id": "req_7e3a1b...",
  "status": "completed",
  "output": "...",
  "agent": "Agent Display Name",
  "duration_ms": 4231
}

Field	Meaning
`task_id`	Identifies the invocation
`request_id`	Correlation ID matching the `request_id` in the original 202 response and `X-Request-Id` on streaming requests
`status`	One of `queued`, `running`, `completed`, `failed`, `timeout`, `canceled`
`output`	Present when `status` is `completed` — expert’s response text
`agent` or `agents_used`	Present when `status` is `completed` — display name(s) of the expert(s) that responded
`duration_ms`	Present when `status` is `completed` or `failed` — wall-clock time from task creation to completion
`error`	Present when `status` is `failed` — error message

Invocation history

GET /api/invocations

curl -s "https://braintrust.ti.trilogy.com/api/invocations?limit=10&agent=archivist" \
  -H "x-api-key: YOUR_API_KEY"

Query parameters:

Param	Description
`limit`	1–100, default 20
`agent`	Filter by agent ID

Response shape: { "invocations": [...] } — each item includes task_id, status, thread_token, mode, input, output, and timestamps.

Traces and Debugging

Every invoke turn returns a UUID request_id (in meta/done, the async 202 body, poll responses, and the X-Request-Id response header). Use it to fetch the LangSmith trace timeline for that turn. Same x-api-key as invoke; you must have hired the expert. GET /api/agents/{agent_id}/traces/{request_id}/timeline Returns run header fields (name, status, start_time, end_time, error) and an interleaved timeline of model and tool events. When truncated is true, the trace exceeded internal fetch limits.

curl -s "https://braintrust.ti.trilogy.com/api/agents/archivist/traces/aaaaaaaa-bbbb-cccc-dddd-eeeeeeeeeeee/timeline" \
  -H "x-api-key: YOUR_API_KEY"

Status	Meaning
`400`	`request_id` is not a valid UUID.
`403`	No expert access, or the run belongs to another user.
`404`	Unknown expert, or no LangSmith run for that `request_id` on this expert.

LangSmith indexing can lag a few seconds after invoke completes — retry if you get 404 immediately after done.

Error Responses

Status	Typical reason	Fix
`400`	Bad JSON, missing `message`, invalid `agent`/`agents`, invalid `attachments` (wrong `s3_key`, extra fields, or more than 20 files), or unsupported extra fields.	Check JSON, `message`, and `agent`/`agents` usage. Verify each attachment’s `s3_key` starts with `invoke-uploads/` and came from the upload-url endpoint. Remove unsupported flags like `async` or `wait_for_completion`.
`401`	Missing or invalid API key	Check the `x-api-key` header name and key value
`403`	Agent not hired or inactive	Hire the agent first; response may include `hire_url`
`404`	Wrong agent ID or task ID	Verify the IDs
`409`	Cancel not allowed for this task state — record exists but is in a terminal state or has no stored internal message id.	Use the cancel endpoint only for `queued` or `running` tasks; check `cancel_applied` and `reason` in the 200 response for diagnostic detail.
`500`	Internal server error	Retry; contact support if repeated
`502`	Conversation backend unavailable	Retry shortly

Windows PowerShell

Use curl.exe — PowerShell’s built-in curl is an alias for Invoke-WebRequest and is not compatible.

curl.exe -s -N -X POST "https://braintrust.ti.trilogy.com/api/agents/invoke" `
  -H "x-api-key: YOUR_API_KEY" `
  -H "Content-Type: application/json" `
  -d '{""agent"":""AGENT_ID"",""message"":""Say hello in one sentence.""}'

Best Practices

Start with streaming mode (-N flag) and parse the initial meta event before any chunk lines.
Move to async mode only when your system needs job-style polling.
Keep agent IDs lowercase with hyphens.
Add context for better, more relevant answers.
Save task_id for traceability and thread_token for conversation continuity.
On multi-turn flows, always send the current thread_token, not a stale one from a prior segment.

Developer

Reference

Base URL

Authentication

Invoke an Expert

Request Body

Attachments

Get a presigned upload URL

Streaming Mode (Default)

Event sequence

Example

Async Mode (`stream: false`)

Step 1 — Create the async task

Step 2 — Poll for the result

Step 3 — Cancel (optional)

Thread Continuity

Multi-turn example

Multi-Agent Collaboration

Delivery Options

Deliver to a Google Chat Space

Deliver as DM from the agent

Delivery diagnostics

Read Endpoints

Task status

Invocation history

Traces and Debugging

Error Responses

Windows PowerShell

Best Practices

Next Steps

Managing API Keys

Triggers

​Base URL

​Authentication

​Invoke an Expert

​Request Body

​Attachments

Get a presigned upload URL

​Streaming Mode (Default)

​Event sequence

​Example

​Async Mode (stream: false)

​Step 1 — Create the async task

​Step 2 — Poll for the result

​Step 3 — Cancel (optional)

​Thread Continuity

​Multi-turn example

​Multi-Agent Collaboration

​Delivery Options

​Deliver to a Google Chat Space

​Deliver as DM from the agent

​Delivery diagnostics

​Read Endpoints

​Task status

​Invocation history

​Traces and Debugging

​Error Responses

​Windows PowerShell

​Best Practices

​Next Steps

Managing API Keys

Triggers

Base URL

Authentication

Invoke an Expert

Request Body

Attachments

Streaming Mode (Default)

Event sequence

Example

Async Mode (`stream: false`)

Step 1 — Create the async task

Step 2 — Poll for the result

Step 3 — Cancel (optional)

Thread Continuity

Multi-turn example

Multi-Agent Collaboration

Delivery Options

Deliver to a Google Chat Space

Deliver as DM from the agent

Delivery diagnostics

Read Endpoints

Task status

Invocation history

Traces and Debugging

Error Responses

Windows PowerShell

Best Practices

Next Steps