OpenAI AgentKit vs Google ADK vs Inngest: Which Agent Platform Should You Choose? (Updated June 2026)

When we first compared these three platforms in late 2025, the question was simple: which one should you build your agents on? A few months into 2026, the question is sharper, because the landscape moved underneath all three. OpenAI announced it is winding down AgentKit's visual Agent Builder, Google reframed its Agent Development Kit from a toolkit into a full execution framework, and Inngest quietly became one of the cleaner ways to run durable agent workflows. This is our updated read on OpenAI AgentKit, Google ADK, and Inngest — what each is now, what changed, and how we'd choose between them today.

All three solve the same underlying problem: how do you build AI systems that do things — schedule, analyze, write, execute — rather than just answer questions? The differences come down to architecture, state and memory, debugging, and how each integrates with the rest of your stack. But the most important 2026 update is about durability of the platform itself, not just the workflows it runs.

The big 2026 change: AgentKit's visual builder is being retired

The headline development since our original comparison: on June 3, 2026, OpenAI announced it is winding down the Agent Builder and Evals products, with availability ending November 30, 2026. The guidance is to move code-based workflows to the Agents SDK and natural-language use cases to Workspace Agents in ChatGPT. If you were planning to build on AgentKit's drag-and-drop canvas, that path now has an expiry date — and the durable, code-first option is the Agents SDK, which received a significant April 2026 overhaul adding native sandboxed execution and a model-native harness for long-running agents across files and tools.

This matters for platform selection in a specific way: the visual-builder convenience that made AgentKit attractive for non-developers is the part going away, while the developer-facing SDK is the part being invested in. We factor that into the recommendations below.

Platform overview (updated June 2026)

Platform	Developer	Primary focus	Best for	2026 status
OpenAI AgentKit / Agents SDK	OpenAI	LLM-powered workflow orchestration	Rapid prototyping, GPT-native agents	Visual Builder retiring Nov 2026; SDK actively expanded
Google ADK	Google	Multi-agent systems, enterprise execution	Data-heavy apps, compliance, observability	Reframed as execution framework; OTel-native
Inngest	Inngest Inc.	Event-driven durable workflows	Background jobs, reliable automation	Ships its own AgentKit framework on top

One naming caution worth clearing up, because it confuses people: Inngest now ships a framework also called "AgentKit" — an orchestration-first layer for building agent networks on top of Inngest's durable execution, with support for OpenAI, Anthropic, and Gemini models and MCP-as-tools. So "AgentKit" in 2026 can mean either OpenAI's product or Inngest's framework. In this comparison, "AgentKit" refers to OpenAI's unless we say otherwise.

Architecture: how they actually work

OpenAI runs on the Responses API, combining chat completions with structured function calling. Conversation history and context live within OpenAI's infrastructure, which simplifies development but means your data sits on OpenAI's servers, and function calls execute synchronously. The 2026 Agents SDK adds sandboxing and a harness for longer-running work, narrowing the historic gap with durable-execution platforms.

Google ADK distributes work across microservices, with agents running in any Google Cloud region and state managed through Firestore, Spanner, or Bigtable. It natively supports multi-agent architectures — specialist agents that collaborate and delegate, with a coordinator routing tasks. The February 27, 2026 integrations expansion is the substantive update: ADK now connects directly to GitHub, Jira, MongoDB, and a set of observability platforms, and it emits OpenTelemetry spans for agent runs, tool calls, and model requests natively. That last point reframes ADK as an OTel-native execution environment — agent behavior flows into the same pipelines that capture your infrastructure metrics, with no custom instrumentation.

Inngest checkpoints every step of a workflow. If a server crashes mid-execution, the workflow resumes exactly where it left off — a design borrowed from durable workflow engines like Temporal and Step Functions. You write standard TypeScript, Python, or Go functions that call any LLM provider, and Inngest handles orchestration, retries with exponential backoff, and state. Wrapping an expensive LLM call in an Inngest function means it doesn't repeat if a later step fails.

Feature comparison

Feature	OpenAI AgentKit / SDK	Google ADK	Inngest
Multi-agent	Handoff orchestration	Hierarchical coordination	Network + router between functions
Memory	Session context (BYO persistence)	Persistent DBs (Firestore/Spanner)	Durable execution state
Observability	Dashboard traces, token usage	OTel-native spans, 5+ obs integrations	Visual replay, step inspection
Setup time	Minutes (SDK)	Hours (GCP setup)	Minutes (CLI)
Models	OpenAI only	Gemini + 100+ via Model Garden	Bring your own (any provider)
Governance	Guardrails, admin controls	VPC, CMEK, IAM policies	Access control, self-host option

The memory and observability rows are where the platforms genuinely diverge. OpenAI keeps memory within its environment — fine for prototyping, limiting for long-term persistence, so you build your own storage layer if you need history beyond a session. Google gives you full control of the store and, as of 2026, first-class telemetry. Inngest's durable state lets a workflow pause for hours or days waiting on an approval, a schedule, or a webhook, then resume without tying up resources.

Real-world deployments

The documented use cases still illustrate each platform's center of gravity. On OpenAI's side, Klarna's support agent handles two-thirds of customer tickets, Clay reported 10x growth with a sales agent, and Ramp built a working buyer agent in hours rather than months. Google uses ADK internally for products like Agentspace and its Customer Engagement Suite, and the platform's compliance posture (HIPAA, FedRAMP, GDPR support) suits regulated industries processing sensitive data at scale. Inngest serves teams whose workflows can't afford to lose state — onboarding pipelines, payment processing, approval chains, scheduled reporting.

The pattern we noted in 2025 has only strengthened: teams combine platforms rather than picking one. A common shape is Inngest as the durable orchestration backbone, with OpenAI's SDK or Google ADK providing the reasoning, so event triggering, state, and error recovery are separated from language understanding. That separation also future-proofs the system — you can swap the AI provider without rewriting the orchestration.

Pricing

Platform	Pricing model	Free tier	Cost character
OpenAI	Pay-per-token + function overhead	Limited	Simple; scales with token volume
Google ADK	API calls + compute + storage	Vertex AI free tier	Complex; optimizable, needs cost discipline
Inngest	Per-event + workflow execution	Developer tier	Predictable; linear with usage

OpenAI's per-token model is the easiest to reason about and the easiest to overrun — long context and many function calls add up fast, so monitor token usage. Google bills compute and storage separately, which lets you optimize inference independently of infrastructure but requires real cost management to avoid surprises. Inngest bills per trigger and execution, with no charge for idle capacity, which keeps background-job costs predictable.

How we'd choose today

Choose OpenAI's Agents SDK when time-to-market matters most, you're committed to GPT models, and you want minimal operational overhead — but build on the SDK, not the retiring visual Builder, if you want the work to last past 2026. Choose Google ADK when data must stay in your infrastructure, you need fine-grained control and compliance, you're already on Google Cloud, or observability is a procurement requirement — its OTel-native telemetry is now a genuine differentiator. Choose Inngest when workflow reliability trumps everything, you want provider flexibility, or you're building background jobs and approval flows where losing state means losing money.

The deeper point hasn't changed: the best platform isn't the objectively most powerful one, it's the one that matches your control requirements, compliance posture, and team's existing stack. What has changed is that platform durability is now part of the decision — a visual builder that's being retired, an execution framework that's investing in observability, and a durable-execution layer that's adding agent primitives are three different bets on where this market is going. Whichever you pick, the reliability of the result depends less on the platform and more on the guardrails you put around it; for that, see our guide to designing safe AI agents, and for the framework-level view, our 2026 framework comparison.

Update, July 2026: OpenAI folds the agent into ChatGPT itself

The comparison above holds, but the landscape moved sharply in July 2026. On July 9, OpenAI launched ChatGPT Work — an agent that lives inside ChatGPT, gathers context across your connected apps and files, and works independently on a goal for hours, returning finished spreadsheets, slides, documents, and web apps. It runs on GPT-5.6 (released the same day in three tiers: Sol, Terra, and Luna), and the Codex app has been merged into a single ChatGPT desktop app carrying Chat, Work, and Codex together.

Two shifts matter for anyone choosing a stack. First, the agent layer is being productized rather than sold as a builder: for a large class of business tasks, the answer is no longer "which framework do I assemble this from" but "do I just turn on the vendor's agent?" Microsoft made the same move, taking its Sales and Service agents in Dynamics 365 to general availability. Second, GPT-5.6's API adds multi-agent capability, running concurrent subagents and synthesizing their work within a single request — pulling orchestration, once a framework's main reason to exist, into the model provider itself.

What this does not settle is governance. A hosted agent that acts across your apps unattended for hours raises exactly the questions this comparison has always ended on: what is it allowed to do, which of its actions are irreversible, and where does a human approve? Each vendor now answers those in its own console and its own format — which is manageable with one agent and unmanageable across three. If you're weighing whether to enable a long-horizon agent for your team, start with the ChatGPT Work governance checklist.