Inngest

The customer's agent tools are run by Inngest rather than supplied by it: its agent tool loop pattern makes each model call and each tool execution a retriable step, and AgentKit gives agents a tool-building API with MCP servers usable as tools and integrations with outside products (E2B, Browserbase, Smithery, Daytona). No connector catalog or authenticated action framework of Inngest's own is documented, so the tools are the customer's code or third-party services.

Sourceinngest.com/docs/ai-patterns/agent-tool-loopsread 2026-09-25

As a workflow engine, Inngest runs functions as steps with retries, parallel steps, fan-out, loops, delays, waits and invocation of other functions, under flow control for concurrency, throttling, debounce, rate limits, batching and priority. For agents, its patterns cover agent tool loops and sub-agent delegation, and AgentKit composes networks of agents with routing.

Sourceinngest.com/docs/learn/inngest-stepsread 2026-09-25

No retrieval structure Inngest maintains over the customer's corpus is documented. Its AI steps and AgentKit run model calls and tools; any indexing, embeddings store or retrieval is the customer's own code or a third-party service.

Sourceinngest.com/docs/features/inngest-functions/steps-workflows/step-ai-orchestrationread 2026-09-25

Durable waits let the customer's code hold for a person: Inngest's human-in-the-loop pattern has a function propose an action, notify someone, and pause on step.waitForEvent (or waitForSignal) until an approval event arrives or a timeout passes. This is a general wait applied to approval, and no approval step, approver identity or review surface of Inngest's own is documented.

Sourceinngest.com/docs/ai-patterns/human-in-the-loopread 2026-09-25

Both compliance and access controls are documented. The security docs state Inngest is SOC 2 Type II compliant with periodic independent security assessments, and HIPAA BAAs are an add-on on Pro and Business and included on Enterprise. For access, SAML authentication is documented for accounts, and the pricing page lists SAML, RBAC and audit trails on Enterprise, with signing keys, event keys and API keys as the platform's credentials.

Sourceinngest.com/docs/learn/securityread 2026-09-25

Every run is recorded for step by step inspection: run details and traces show each step, AI calls through step.ai carry prompts, tokens and responses, the AI Overview dashboard turns that metadata into usage, cost and performance views that click through to the runs behind them, and Insights queries runs and events. Trace history runs by plan (24 hours on Free to 90 days on Enterprise), with trace and log exports and Datadog and Prometheus integrations.

Sourceinngest.com/docs/platform/monitor/ai-overviewread 2026-09-25

Every step is checkpointed, so an agent's progress is a durable record and a failed run resumes without repeating finished steps; that is state inside one run, and the code carries on from it. Across runs, AgentKit's History Adapter keeps conversations per thread so a later run resumes in context. That store is the customer's own database behind an adapter the customer writes, with no published lifetime or purge path and no first-party surface to review or edit it, and AgentKit's long-term memory runs on Mem0, a third party.

Sourceinngest.com/docs/learn/durable-agentsread 2026-09-25

Inngest deploys into the customer's own environment two ways: functions always run on the customer's own hosts (serverless platforms, containers or servers) and Inngest calls them, and the Inngest server itself self-hosts under a documented guide since its 1.0 release. The server is source available under the SSPL, with each release becoming Apache 2.0 on its third anniversary (per the repository's license file).

Sourceinngest.com/docs/self-hostingread 2026-09-25

Ready made starting points are published: AgentKit examples and a starter for a conversational app with persistent history, AI patterns in the docs (agent tool loops, human-in-the-loop, sub-agent delegation), and agent skills and plugins for coding agents. These are examples a developer builds from, not a catalog of agents a buyer adopts.

Sourceagentkit.inngest.com/examples/overviewread 2026-09-25

Functions start on their own: they trigger on events (including webhook events Inngest consumes), on cron schedules, and on several triggers at once, and a running function can wait for an event or signal before carrying on.

Sourceinngest.com/docs/guides/scheduled-functionsread 2026-09-25

The customer chooses the model through a surface of Inngest's own. step.ai.infer offloads a model call to Inngest's infrastructure, with the model named through Inngest's provider adapters (for example step.ai.models.openai with a chosen model), step.ai.wrap wraps the OpenAI, Anthropic and Vercel AI SDKs as steps, and AgentKit takes a model per agent. All of them track requests and responses in traces.

Sourceinngest.com/docs/features/inngest-functions/steps-workflows/step-ai-orchestrationread 2026-09-25

Inngest is callable from outside through SDKs in TypeScript, Python and Go, a REST API, and the CLI, and it provides MCP servers for Inngest Cloud and the local Dev Server whose tools match the REST API and CLI commands, so a coding agent can inspect and operate apps, functions, events and runs.

Sourceinngest.com/docs/ai-dev-tools/mcpread 2026-09-25

Evaluation of agent output is documented. Agent Evals attach numeric or boolean quality scores to runs and steps, including LLM-as-a-judge scores and deferred scoring, group related work into sessions, and compare prompts, models or strategies with group.experiment() A/B variants. The docs label scoring beta, a public beta customers can use today. Scores are queryable in Insights and metered by plan (10,000 on Free to 250,000 on Business).

Sourceinngest.com/docs/learn/agent-evalsread 2026-09-25

No browser or desktop the agent operates, and no browser environment of Inngest's own, is documented. AgentKit integrates with Browserbase, but the browser and its control belong to Browserbase, a third party, and Inngest runs the customer's function around it.

Sourceagentkit.inngest.com/integrations/browserbaseread 2026-09-25