Amplitude AI agents: APIs, apps, and governed workflows
Understand Amplitude's native agents and verified APIs, then see how Major adds deterministic apps, durable state, and approval controls to analytics-driven operations.

Amplitude AI agents analyze product behavior and can act through Amplitude's connected tools. This guide covers their native capabilities, verified APIs, and how Major adds deterministic apps, durable process state, and approval controls around analytics-driven work.
For customer success and product operations leaders, the practical question is how a usage signal becomes a reviewed action without losing its evidence or creating duplicate work.
What Amplitude does for product and customer teams
Amplitude organizes behavioral data around events, users, projects, and analyses. Product managers investigate activation funnels and retention, analysts compare segments, and growth teams examine conversion. Its Dashboard REST API exposes results from these analyses rather than requiring someone to copy numbers from a chart.
Customer success operations has a different job: connect those usage patterns to an account owner and a follow-up process. A quiet account might be at risk, observing a holiday, or missing instrumentation. A decline alone cannot distinguish those explanations.
Major is the enterprise platform where agents build the software they run on. Here, an agent can build a deterministic usage-review app, then reuse its calculations, account mappings, and history on subsequent runs. The app holds state; the agent assesses ambiguous evidence; the workflow coordinates collection, routing, and approvals.
That division matters when several teams share the same signal. Product needs a reproducible definition of activation, while customer success needs a durable explanation of why an account entered its queue. Both should be able to inspect the underlying window and metric.
Amplitude's built-in AI agents
Amplitude already offers substantial AI functionality. Its AI Agents product page describes Global Agent as an analyst that understands product data, analyzes experiments, creates dashboards, segments users, and maintains taxonomy. Custom Agents support recurring jobs that run on a schedule or on demand with configured instructions and tool access.
The same page describes capabilities for deep analysis, observing session friction, and conversion experimentation. Agents can generate charts, dashboards, notebooks, and draft product documents. Native agent connectors bring in external context and send updates to connected tools. It would be inaccurate to claim Amplitude only answers questions or cannot work across systems.
Major adds a different operating layer when the process needs its own application: account crosswalks, review screens, persistent action records, approval deadlines, and deterministic recovery after failures. An agent on Major still makes the judgment call, such as whether a usage drop warrants attention.
Choose the boundary around ownership. Keep behavioral analysis in Amplitude. Put the shared operational record in a Major app when customer success, product, and IT need to govern the same process across runs. Native Amplitude agents and agents on Major are distinct, complementary capabilities.
How to connect Amplitude to Major
Amplitude has no native or Pipedream connector in the catalog: it connects as a custom API connector, with base URL and authentication configured once by IT. Major orchestrates Amplitude; it never replaces it.
Start with the API family required by the job. Dashboard REST and Export use https://amplitude.com by default or https://analytics.eu.amplitude.com for EU projects. The Developer API instead uses https://developer-api.amplitude.com or https://developer-api.eu.amplitude.com. Do not send one family's credentials indiscriminately to another host.
Build the operating pieces on Major:
- IT configures the custom connector and stores credentials in the central vault, never in code or prompts.
- The app stores account mappings, completed data windows, candidate signals, and review decisions in managed database tables.
- The agent receives scoped access to evidence and recommendation fields, not unrestricted external writes.
- The workflow schedules collection, invokes judgment, routes approval, and records confirmed outcomes.
SSO, RBAC, and audit logs cover every app, agent, and workflow. Connector operation restrictions are especially important when a vendor credential itself is broader than the task.
Amplitude API guide for operational builds
Authentication and scope boundaries
API Authentication specifies HTTP Basic authentication for Dashboard REST and Export: project API key as username, secret key as password. These are project credentials, not OAuth scopes. The keys guide distinguishes public ingestion keys, private project secrets, Data tokens, Experiment keys, and organization-level credentials. Do not request organization-wide credentials for one project's usage review.
The Developer API overview documents OAuth device-flow access tokens and long-lived personal access tokens, both sent as Bearer credentials. Relevant scopes include projects:read and analytics:read. Its taxonomy:read scope permits taxonomy inspection. Those scopes do not turn a Dashboard REST project secret into a read-only credential.
Pagination, limits, and test options
Developer API lists use cursor and limit parameters, defaulting to 50 items with a maximum of 200. Follow the next cursor in the pagination object until no further page remains. Dashboard user activity instead uses offset and limit, with up to 1,000 requested events; responses can exceed that to avoid splitting sessions. Export uses inclusive start and end upload-hour windows, not cursor pagination.
Dashboard REST limits include five concurrent requests per API key, plus query-cost budgets of 1,000 concurrent cost and 108,000 cost per hour. User activity and user search additionally share 360 queries per hour. Export defaults are four concurrent requests and 360 requests per hour per project key. Account overrides can apply.
For Dashboard throttling, watch HTTP 429 and honor Retry-After when present. Numerical Developer API limits and a standard remaining-budget header are not documented publicly in the overview reviewed here. Its error format can indicate retryability and a suggested delay. Amplitude recommends a separate test project or development environment for each production project.
Events and webhook delivery
Webhooks Streaming forwards selected events when ingested, and user updates when events or Identify calls arrive. It is not a scheduled chart-change alert. The integration supports five extra headers; configure a secret header and validate it before accepting requests. A native cryptographic signing scheme is not documented on that page.
Delivery failures receive nine further attempts over four hours. Within an attempt, 5xx and 429 responses can trigger up to three immediate attempts with exponential backoff and jitter, bounded by four seconds. Design the receiver for repeated delivery, not exactly-once execution.
GET /v1/context
This Developer API endpoint returns the authenticated principal, granted scopes, and organization context.
- App: Show the verified organization beside a connection configuration.
- Agent: Check whether available access supports the requested analysis.
- Workflow: Validate the connection before a scheduled processing run.
GET /api/3/chart/:chart_id/query
Read JSON results from a saved chart using credentials for its owning project.
- App: Store the chart identifier and retrieved evidence with a review case.
- Agent: Interpret changes against the approved metric definition.
- Workflow: Refresh the chosen chart before generating the weekly account review.
GET /api/2/events/list
List event types so a build can refer to actual instrumentation.
- App: Populate an event picker for an operations owner.
- Agent: Flag unclear naming before interpreting a candidate signal.
- Workflow: Check that a configured event remains available before querying it.
GET /api/2/events/segmentation
Query event trends with specified dates, event definitions, and segmentation.
- App: Persist comparable usage summaries for tracked accounts.
- Agent: Assess whether the change reflects adoption loss or a segment shift.
- Workflow: Fetch equivalent periods and run deterministic candidate checks.
GET /api/2/funnels
Read conversion through a defined sequence of events.
- App: Display an activation case with its affected funnel stage.
- Agent: Prioritize the stage most likely to explain the reported friction.
- Workflow: Refresh the same funnel definition after a release review.
GET /api/2/retention
Query return behavior relative to starting and returning events.
- App: Retain a dated retention snapshot beside the cohort definition.
- Agent: Explain whether weaker repeat use merits investigation.
- Workflow: Compare only sufficiently mature retention windows before routing cases.
GET /api/2/useractivity
Retrieve a user summary and recent or earliest activity using an Amplitude user identifier.
- App: Show a permissioned evidence panel for an assigned investigation.
- Agent: Summarize the relevant sequence without exposing unrelated personal activity.
- Workflow: Retrieve a bounded event sample only after a case qualifies.
GET /api/3/cohorts
List discoverable cohorts and their definitions, rather than assuming every list contains membership rows.
- App: Let the owner select an existing cohort for monitoring.
- Agent: Assess whether that cohort fits the question being investigated.
- Workflow: Refresh metadata and detect changed monitoring definitions.
GET /api/2/export
Export raw events as a ZIP archive of gzip-compressed, line-delimited JSON files.
- App: Store validated account aggregates and completed upload-hour checkpoints.
- Agent: Evaluate the resulting evidence without reading an entire archive into context.
- Workflow: Download, validate, aggregate, and advance checkpoints only after success.
Amplitude automation use cases on Major
Usage decline to a Gainsight review
This is grounded in the published Amplitude and Gainsight process linked in the library below. Compare meaningful activity against the account's own baseline, then consider seasonality and previous interventions.
- App: Account crosswalk, usage history, review queue, and existing CTA references.
- Agent: Judge whether the decline merits attention and draft its rationale.
- Workflow: Schedule GET /api/2/export, aggregate complete windows, resolve the Gainsight account, and create a CTA only after approval.
Activation friction to a Linear investigation
This proposed pattern connects product analysis to an owned investigation instead of opening an issue for every fluctuation. Keep metric changes separate from confirmed defects.
- App: Activation cases, linked release notes, owners, and issue references.
- Agent: Interpret funnel evidence and draft a focused investigation brief.
- Workflow: Refresh GET /api/2/funnels, collect permitted release context, and route the draft before creating a Linear issue.
Retention review with Salesforce and Slack
This proposed pattern combines a mature behavioral window with account context. It should produce an internal recommendation, not an automatic customer message or an unsupported churn prediction.
- App: Account mappings, retention snapshots, review decisions, and suppression periods.
- Agent: Assess whether the usage pattern and commercial context justify outreach.
- Workflow: Read GET /api/2/retention, retrieve permitted Salesforce context, and request approval through Slack before a downstream update.
The Gainsight, Linear, and Salesforce hubs cover those systems' roles. The examples above describe operating patterns, not prebuilt integrations or guaranteed connector operations.
Architecture: build the state before adding judgment

Figure 1. The app holds state while the agent assesses risk and the workflow coordinates action.
Begin with a scheduled closed-window pull. A single behavioral event can arrive quickly, but a usage-decline assessment needs a complete comparison period. The workflow retrieves data through the configured connector and calls deterministic validation and aggregation code.
The Major app holds the account crosswalk, window status, computed metrics, evidence references, and review queue. The agent receives that bounded context and makes the judgment: meaningful risk, explainable variation, or insufficient evidence. The workflow then routes the recommendation according to policy.
Keep the external action separate from the recommendation. A reviewer sees the same numbers and definition the agent saw, plus the intended destination and owner. An approved action writes its external reference back to the app, so the next run can find it.
When an agent figures out a repeatable collection or validation step, that step can become app code instead of fresh reasoning on every run. The agent still interprets uncertainty. Reason once. Run forever applies to the reusable software, not to pretending every future judgment is settled.
A worked example with numbers
For example, suppose a customer success operations team reviews 200 accounts each week. This is an illustrative planning scenario, not a customer result or a promised saving.
Manual preparation takes 12 minutes per account: retrieve usage, compare periods, find the owner, and prepare a review note. That is 200 × 12 = 2,400 minutes, or 40 hours per week.
In the proposed process, the workflow collects closed windows and the app calculates comparisons. Suppose 30 accounts need human attention at eight minutes each, while the other 170 receive a one-minute quality spot-check. Review time is 240 + 170 = 410 minutes. Add 70 minutes for operations checks, and the total is 480 minutes, or eight hours. The illustrative difference is 32 hours, excluding initial setup and ongoing changes.
One account falls from 1,000 meaningful events to 600, a 40% decline. The workflow verifies complete windows and an exact account match. The agent considers a planned shutdown and recommends suppression rather than outreach. The reviewer records that reason, preserving context for next week's run.
Governance, approval, and failure handling

Figure 2. Incomplete evidence goes to investigation; approved recommendations become recorded actions.
Let validated reads, calculations, and draft creation proceed automatically. Require a person before consequential writes, customer outreach, taxonomy changes, or experiment changes. The agent must never approve its own recommendation, invent missing evidence, or interpret a failed pull as zero usage.
Use a stable case key combining project, account, metric, and window. Store recommendation status, reviewer identity, approval expiry, and the external action reference. After a timeout on an external write, reconcile the result before retrying; do not assume an identifier guarantees vendor-side idempotency.
Export documentation says data can become available within two hours after receipt, uses UTC upload hours, and limits responses to 6 GB. Validate archives before advancing checkpoints. Missing recent files mean incomplete evidence, not confirmed inactivity.
Coordinate rate budgets across workers, honor documented retry guidance, and quarantine invalid credentials or parameters rather than retrying indefinitely. Retain audit records of reads, judgments, approvals, and outcomes. Apply retention and access policies to behavioral evidence, especially user-level activity, rather than copying every available event into the review app.
Amplitude AI Apps, Agents and Workflow Automations
The current library contains one dedicated Amplitude implementation guide. Start there if customer success owns the process and Gainsight is the destination:
- Turn Amplitude usage declines into reviewed Gainsight churn-risk CTAs: This guide covers account-level baselines, evidence-backed risk assessment, CSM approval, and durable CTA references. It explains how to keep incomplete exports and repeated signals from becoming misleading follow-up tasks.
Use this hub to choose API surfaces and operating boundaries; use the linked guide for the detailed build. Its subject is usage-risk review, not an all-purpose prediction engine. The distinction matters when a team wants to copy a threshold without agreeing on meaningful events or account mapping.
Before adapting the guide, name the owner of the metric and the person authorized to approve follow-up. Record the account identifier, comparison window, and definition of meaningful usage. Those choices determine whether a decline can be reviewed consistently. The other proposed patterns in this hub are operating examples, not published implementation guides.
Frequently asked questions
What are Amplitude AI agents?
Amplitude's native agents analyze product data, create artifacts, and perform supported actions through connected tools. Agents on Major can use Amplitude evidence within a separate governed process whose deterministic app holds state, approvals, and action history.
Does Amplitude have its own AI features?
Yes. Amplitude documents Global Agent, Custom Agents, and capabilities for analysis, session observation, and conversion experimentation. Its native agent connectors also support cross-tool context and actions, so the distinction is not simply analytics versus automation.
How does Amplitude connect to Major?
Amplitude connects through a custom API connector, not a native or Pipedream connector in the current catalog. IT configures the regional base URL and authentication once, with credentials stored centrally and access restricted to approved operations.
Which authentication does the Amplitude API use?
Dashboard REST and Export use HTTP Basic authentication with a project API key and secret key. Developer API requests generally use an OAuth access token or personal access token as a Bearer credential. HTTP V2 ingestion instead takes a project API key in the request body.
Can Amplitude webhooks trigger a workflow on Major?
Yes, with a configured receiving integration. Webhooks Streaming forwards selected events as Amplitude ingests them and can forward user updates. Configure a secret header and validate it at the receiver. The referenced documentation does not specify a native cryptographic signature, and repeated deliveries must be deduplicated.
Can an agent turn declining usage into a Gainsight alert?
Yes. A workflow can collect complete usage windows, aggregate by account, and ask an agent to assess the decline alongside account context. A Major app holds the evidence and review queue, and a person approves the Gainsight CTA before the workflow creates it.
Is there an Amplitude sandbox for testing automation?
Amplitude recommends creating a test project or development environment for each production project. Use isolated credentials and synthetic activity, then test incomplete exports, duplicate events, throttling, and rejected approvals before enabling production writes.
Sources
- Amplitude Developer API overview: regional hosts, authentication, scopes, list pagination, errors, idempotency, and dry runs.
- Get authenticated context: the principal, organization, authentication type, and granted scopes returned by the context endpoint.
- API Authentication: project API key and secret key usage for Analytics REST requests and ingestion differences.
- Keys and Tokens: project credentials, Data tokens, Experiment keys, and organization-level key boundaries.
- Dashboard REST API: saved charts, event lists, segmentation, funnels, retention, user activity, and throttling guidance.
- Export API: raw event archives, inclusive upload-hour ranges, availability delay, size limits, and request budgets.
- Behavioral Cohorts API: discoverable cohorts, cohort definitions, asynchronous downloads, and documented download restrictions.
- HTTP API v2: event ingestion, request authentication, throttling responses, and event deduplication.
- Webhooks Streaming: event and user forwarding, custom headers, payload configuration, and delivery retry behavior.
- Create a project in Amplitude: project isolation, separate API keys, and the recommendation to maintain test environments.
- Manage organizations and projects: project administration, access, keys, and the effect of project time-zone settings.
- Amplitude AI Agents: Global Agent, Custom Agents, analysis, observation, experimentation, artifacts, and native agent connectors.
Start building on Major
Build an Amplitude review app that keeps usage evidence, agent judgment, and approved actions in one governed process. Start with one metric and one accountable reviewer.
About the author
Ritvij Gautam is Head of Growth at Major.
Last reviewed November 20, 2026 by Ritvij Gautam.
Frequently asked questions
- What are Amplitude AI agents?
- Amplitude's native agents analyze product data, create artifacts, and perform supported actions through connected tools. Agents on Major can use Amplitude evidence within a separate governed process whose deterministic app holds state, approvals, and action history.
- Does Amplitude have its own AI features?
- Yes. Amplitude documents Global Agent, Custom Agents, and capabilities for analysis, session observation, and conversion experimentation. Its native agent connectors also support cross-tool context and actions, so the distinction is not simply analytics versus automation.
- How does Amplitude connect to Major?
- Amplitude connects through a custom API connector, not a native or Pipedream connector in the current catalog. IT configures the regional base URL and authentication once, with credentials stored centrally and access restricted to approved operations.
- Which authentication does the Amplitude API use?
- Dashboard REST and Export use HTTP Basic authentication with a project API key and secret key. Developer API requests generally use an OAuth access token or personal access token as a Bearer credential. HTTP V2 ingestion instead takes a project API key in the request body.
- Can Amplitude webhooks trigger a workflow on Major?
- Yes, with a configured receiving integration. Webhooks Streaming forwards selected events as Amplitude ingests them and can forward user updates. Configure a secret header and validate it at the receiver. The referenced documentation does not specify a native cryptographic signature, and repeated deliveries must be deduplicated.
- Can an agent turn declining usage into a Gainsight alert?
- Yes. A workflow can collect complete usage windows, aggregate by account, and ask an agent to assess the decline alongside account context. A Major app holds the evidence and review queue, and a person approves the Gainsight CTA before the workflow creates it.
- Is there an Amplitude sandbox for testing automation?
- Amplitude recommends creating a test project or development environment for each production project. Use isolated credentials and synthetic activity, then test incomplete exports, duplicate events, throttling, and rejected approvals before enabling production writes.