By R. Shivakumar (rshivakumar@protonmail.com)
Three platforms are fighting to become the standard for building AI agents: OpenAI's AgentKit, Google's Agent Development Kit (ADK), and Inngest. Each takes a different approach to solving the same problem—how do you build AI systems that actually do things instead of just answering questions?
An Agent Development Kit gives developers the tools to build autonomous systems powered by large language models. Unlike chatbots that simply respond to queries, agents built with these platforms can execute multi-step tasks: scheduling meetings, analyzing data, writing reports, or running code. The key differences come down to architecture, memory management, debugging capabilities, and how they integrate with other systems.
OpenAI launched AgentKit on October 6, 2025 with the promise of taking agents from prototype to production faster. Google announced ADK at Cloud NEXT 2025 targeting enterprise customers who need fine-grained control and security. Inngest, originally built for reliable workflow automation, has become popular with teams that want durable execution and don't mind bringing their own LLM provider.
Platform Overview
| Platform | Developer | Primary Focus | Best For |
|---|---|---|---|
| OpenAI AgentKit | OpenAI | LLM-powered workflow orchestration | Rapid prototyping, GPT integration |
| Google ADK | Google DeepMind | Multi-agent systems, enterprise scale | Data-heavy apps, strict compliance |
| Inngest | Inngest Inc. | Event-driven durable workflows | Background jobs, reliable automation |
OpenAI AgentKit builds on the Responses API released in March 2025. The platform centers on Agent Builder, a visual canvas where developers design workflows with drag-and-drop nodes. According to OpenAI, partners like Ramp went from blank canvas to working agent in "just a couple of hours." Klarna built a support agent that now handles two-thirds of customer service tickets.
Google ADK gives enterprises the control they demand. The open-source framework works with any LLM through Vertex AI Model Garden and LiteLLM integration, though it's optimized for Gemini. ADK is the same framework Google uses internally for products like Agentspace and their Customer Engagement Suite. Teams can deploy agents as microservices across Google Cloud Platform with full access to enterprise databases.
Inngest takes a different path. The platform focuses on durable execution—workflows that survive crashes, continue after hours or days, and automatically retry failed steps. While Inngest doesn't include AI reasoning out of the box, developers write standard TypeScript, Python, or Go functions that call any LLM provider, with Inngest handling orchestration and state management.
Architecture: How They Actually Work
OpenAI AgentKit runs on the Responses API, which combines chat completions with structured function calling. The architecture keeps conversation history and context within OpenAI's infrastructure. This simplifies development but means your data lives on OpenAI's servers. Function calls execute synchronously—the agent waits for each step to complete before moving forward.
Google ADK distributes work across microservices. Agents can run on any Google Cloud region, with state managed through Firestore, Spanner, or Bigtable depending on your needs. The platform supports both sync and async execution patterns. Multiple specialized agents can work together on complex problems, with one coordinator agent delegating tasks to sub-agents focused on specific capabilities.
Inngest checkpoints every step of a workflow. If your server crashes mid-execution, the workflow picks up exactly where it left off. This design comes from workflow engines like Temporal and AWS Step Functions. When you wrap LLM API calls in Inngest functions, expensive AI operations don't need to repeat if downstream steps fail.
Key Features Compared
| Feature | OpenAI AgentKit | Google ADK | Inngest |
|---|---|---|---|
| Integration | Responses API, built-in tools | Vertex AI, Model Garden, LiteLLM | HTTP, webhooks, event triggers |
| Multi-Agent | Chain-of-thought orchestration | Hierarchical agent coordination | Workflow routing between functions |
| Memory | Session-based context | Persistent databases (Firestore/Spanner) | Durable execution state |
| Debugging | Dashboard traces, step recordings | Cloud Logging & Monitoring | Visual replay, step-by-step inspection |
| Setup Time | Minutes with SDK or visual builder | Hours (requires GCP setup) | Minutes with CLI |
| Connectors | MCP, pre-built integrations | Native Google services, 100+ enterprise | HTTP endpoints, cloud platforms |
| Models | GPT-4o, o-series reasoning models* | Gemini, 100+ via Model Garden | Bring your own (any provider) |
| Governance | Guardrails, admin controls | VPC, CMEK, IAM policies | Access control, permissions |
*Model note: GPT-4.5 was the flagship model at AgentKit's launch but was deprecated from the API on July 14, 2025. It remains available in ChatGPT Pro. Developers should use GPT-4.1, GPT-4o, or o-series models via the API.
Memory management separates these platforms. OpenAI handles memory within its environment—great for prototyping, limiting for long-term persistence. If you need conversation history beyond the current session, you'll build your own storage layer.
Google ADK gives you full control. Pick Firestore for document storage, Spanner for globally consistent data, or Bigtable for high-throughput workloads. This matters for enterprises with audit requirements, compliance mandates, or data retention policies.
Inngest's durable state lets workflows pause indefinitely and resume seamlessly. A workflow can wait hours or days for external events—user approvals, scheduled times, webhook callbacks—without tying up resources. The platform handles exponential backoff retries automatically.
Debugging tools vary dramatically. Inngest provides a visual interface where you replay failed runs step-by-step, inspect inputs and outputs at each stage, and modify state to test fixes. You can see exactly what went wrong and why.
OpenAI's dashboard shows trace logs with reasoning chains, function calls, and token usage. This helps you understand agent decision-making and optimize prompts.
Google ADK integrates with Cloud Logging and Monitoring. You get advanced capabilities like custom metrics, distributed tracing, and alerting. The trade-off? A steep learning curve if you're new to Google Cloud.
For integrations, Google ADK wins for existing GCP customers. Agents access Sheets, Docs, Gmail, and Calendar without complex auth flows. AgentKit supports external APIs through the Model Context Protocol (MCP), letting agents interact with third-party services through standardized interfaces. Inngest excels at event-driven patterns with native support for HTTP webhooks, cron schedules, and cloud platform events.
Real-World Deployments
The documented use cases reveal each platform's strengths.
OpenAI AgentKit moved fast for early adopters. According to OpenAI's case studies, Klarna built a support agent handling two-thirds of customer tickets. Clay scaled sales agents and saw 10x growth. Ramp's team built a buyer agent that went from concept to working prototype in hours using Agent Builder's visual canvas. The agent now helps with procurement decisions, integrating with their internal systems through custom functions.
At DevDay 2025, an OpenAI engineer demonstrated building a complete agent in under 8 minutes on stage. The speed advantage is real.
Google ADK hasn't published specific customer case studies yet—the platform is newer and targets enterprise customers who typically don't publicize internal tools. However, Google uses ADK internally for products like Agentspace and their Customer Engagement Suite, demonstrating production readiness.
The platform's architecture suits data-intensive scenarios. With HIPAA-compliant infrastructure, support for healthcare data formats, and deep integration with BigQuery, ADK makes sense for regulated industries processing sensitive information at scale.
Inngest serves teams building production workflows that can't afford downtime. The durable execution model works well for customer onboarding pipelines, payment processing, complex approval chains, and scheduled reporting systems. When your workflow spans multiple steps with different failure modes—API calls, database writes, external webhooks—Inngest ensures nothing gets lost.
Teams typically combine platforms. An e-commerce company might use AgentKit for a customer service chatbot, Google ADK to analyze purchase patterns and inventory data, and Inngest to orchestrate the order fulfillment pipeline. This hybrid approach uses each platform's strengths without vendor lock-in.
Developer Experience: What It's Actually Like
OpenAI AgentKit prioritizes speed. Developers can build functional agents in minutes using either code or the visual builder. The Agents SDK abstracts complexity around prompt engineering, function calling, and context management. Early reports suggest significant time savings—Ramp's procurement agent took hours instead of months.
The main risk? Poorly configured connectors can expose sensitive data. Since agents make autonomous decisions about which functions to call, you need thoughtful permission design. OpenAI provides guardrails to prevent PII leaks and block jailbreak attempts, but security remains the developer's responsibility.
Google ADK requires Google Cloud expertise. You'll need to understand projects, service accounts, IAM roles, and VPCs before you start. The payoff is complete control over agent behavior, model selection, inference parameters, and deployment topology.
Advanced features like trajectory evaluation—comparing what an agent did versus what it should have done—help teams systematically improve performance. Combined with A/B testing frameworks, you can data-drive agent optimization instead of guessing.
Inngest emphasizes local development. The dev server provides real-time visual debugging on your machine before deploying to production. You see every function call, state transition, and API request. The visual debugger makes it straightforward to identify bottlenecks and understand failure modes.
Since Inngest doesn't include AI reasoning, you control the LLM provider relationship. Write functions that call OpenAI, Anthropic, Google, or any other provider, with Inngest handling only orchestration. This separation appeals to teams that want flexibility as the AI landscape evolves.
Security and Compliance
Security architecture differs fundamentally.
OpenAI AgentKit processes data in OpenAI's infrastructure. You're trusting OpenAI's security practices. The company is SOC 2 Type II certified and supports data residency requirements for enterprise customers. For most consumer applications, this trade-off is acceptable. Regulated industries may want more control.
Google ADK lets organizations keep data within their own cloud projects. Configure VPC Service Controls to prevent data from leaving designated networks. Use Customer-Managed Encryption Keys (CMEK) for data at rest. Implement detailed audit logging. The platform supports HIPAA, FedRAMP, and GDPR compliance. Government agencies and enterprises with strict data sovereignty requirements often need these capabilities.
Inngest operates as an orchestration layer. The platform records workflow metadata—timestamps, success/failure status, execution paths—but business data flows through your application's functions. Deploy Inngest in self-hosted mode for complete control, or use the cloud version with appropriate access controls. This architecture offers a middle ground between convenience and control.
Pricing: What You'll Actually Pay
| Platform | Pricing Model | Free Tier | Details |
|---|---|---|---|
| OpenAI AgentKit | Pay-per-token + function overhead | Limited usage | API Pricing |
| Google ADK | API calls + compute + storage | Vertex AI free tier | Vertex Pricing |
| Inngest | Per-event + workflow execution | Developer tier | Inngest Pricing |
OpenAI's model is simple but scales with usage. You pay for input and output tokens plus function call overhead. Short interactions remain affordable. Long documents, extensive conversation history, or numerous function calls add up quickly. Monitor token usage and optimize prompts to control costs.
Google ADK charges for compute and storage separately, which integrates with existing Vertex credits. The pricing model is complex—you can optimize inference costs independently from infrastructure costs. Running agents on preemptible compute instances cuts costs significantly for batch workloads. The complexity requires cost management expertise to avoid surprises.
Inngest bills per trigger and workflow execution. Costs are predictable for background tasks. You pay for what you use with no charges for idle capacity. The free tier is generous for side projects and small applications. As usage scales, costs remain linear and predictable.
Choosing Your Platform
Pick OpenAI AgentKit when:
- Time-to-market matters more than infrastructure control
- You're building conversational interfaces or content generation tools
- The GPT ecosystem fits your use case
- You want minimal operational overhead
- Your team lacks deep infrastructure expertise
AgentKit works well for customer service bots, sales automation, content generation, and intelligent assistants focused on natural language understanding.
Pick Google ADK when:
- You're processing sensitive data that must stay in your infrastructure
- You need fine-grained control over agent behavior
- Your use case requires multimodal capabilities (text, images, video, audio)
- You're already using Google Cloud services
- Compliance and auditability are critical
- You're building complex multi-agent systems
ADK fits analytics applications, document processing systems, and scenarios requiring coordination between specialized agents.
Pick Inngest when:
- Workflow reliability trumps everything else
- You need guaranteed execution despite failures
- Workflows span multiple steps with different failure modes
- You want flexibility in AI provider relationships
- You're building background jobs, scheduled tasks, or webhook-driven automation
- Cost predictability matters
Inngest excels for payment processing, customer onboarding, scheduled reports, approval workflows, and any scenario where losing state means losing money.
Many teams combine platforms: AgentKit for user-facing interactions, Google ADK for data analysis, Inngest for backend reliability. This hybrid approach avoids vendor lock-in while leveraging each platform's strengths.
Performance Reality
OpenAI AgentKit delivers 1-3 second response times for simple queries and 5-10 seconds for complex reasoning with multiple function calls. The platform handles concurrency well—OpenAI's infrastructure autoscales. Rate limits apply, especially for lower-tier accounts, which can constrain high-volume applications.
Google ADK with TPU acceleration handles large datasets and parallel agent execution. The platform processes thousands of requests per second when configured properly. Autoscaling policies adjust compute resources based on demand. Batch processing enables efficient handling of large-scale jobs—analyzing millions of documents, generating reports across entire datasets.
Inngest specializes in concurrency, processing thousands of triggers with sub-second latency. The architecture handles traffic spikes gracefully, queuing events and processing them without dropping requests. Durable state ensures long-running tasks continue uninterrupted even for workflows spanning hours or days. For unpredictable traffic patterns, Inngest's elastic scaling prevents over-provisioning while maintaining consistent performance.
Community and Ecosystem
OpenAI has built a large, active community with extensive documentation and third-party resources. The Model Context Protocol is gaining adoption with numerous integrations from both OpenAI and the community. Developer forums and Discord channels provide peer support.
Google ADK benefits from Google Cloud's extensive partner network. Enterprises can work with system integrators experienced in GCP deployments. The platform integrates seamlessly with other Google services. The community is smaller, reflecting ADK's enterprise focus and recent launch.
Inngest has cultivated a developer-friendly community focused on reliability. The platform's open-source roots create alignment between the company and users. Community contributions include SDKs for additional languages, integration templates, and deployment guides.
Migration and Integration
Moving between platforms takes planning. Migrating from AgentKit to ADK means rewriting agent logic for Vertex AI APIs and Gemini models. Core reasoning chains translate relatively easily, but you'll rebuild state management, integrate with Google Cloud services, and adapt deployment patterns.
Adding Inngest to existing workflows is straightforward. Wrap existing LLM API calls in Inngest functions to gain reliability and observability without fundamental changes. Teams adopt durable execution for critical workflows while maintaining existing agent implementations.
The emerging pattern: use Inngest as the orchestration backbone with OpenAI or Google ADK providing AI reasoning. Inngest handles event triggering, state management, and error recovery. The AI platform focuses on language understanding and decision-making. This separation creates more maintainable systems and provides flexibility to swap AI providers as technology evolves.
Trade-Offs
| Platform | Strengths | Limitations |
|---|---|---|
| OpenAI AgentKit | Fast setup, strong reasoning, growing ecosystem | Vendor lock-in, scaling costs, data in OpenAI's hands |
| Google ADK | Enterprise-grade, multimodal, full control | Complex setup, steep learning curve, higher infrastructure costs |
| Inngest | Reliable execution, affordable, flexible providers | No built-in AI, smaller ecosystem, requires LLM integration |
What's Next
OpenAI is investing in agent capabilities with continuous model improvements. Expect deeper enterprise integrations and better cost management tools.
Google ADK is extending multimodal capabilities and simplifying the developer experience. The platform is adding pre-built templates for common use cases and streamlining deployment.
Inngest is developing higher-level abstractions for AI patterns like retrieval-augmented generation and agent orchestration while maintaining its core reliability focus.
The future likely involves hybrid setups. Teams will use OpenAI for reasoning, Google ADK for data processing, and Inngest for reliability. As the ecosystem matures, cross-platform compatibility and open standards may let agents move seamlessly between systems.
Bottom Line
AI agents have moved beyond experiments—they're becoming core infrastructure. Your choice shapes workflow efficiency and scalability for years.
OpenAI AgentKit delivers speed for teams prioritizing rapid iteration. Google ADK provides enterprise capabilities for organizations with complex requirements. Inngest ensures reliability for teams building production systems that must handle failures gracefully.
The best choice isn't about which platform is objectively superior. It's about which aligns with your technical capabilities, compliance requirements, and strategic direction.
About the author: R. Shivakumar is an independent technology analyst specializing in AI infrastructure and developer platforms. Contact: rshivakumar@protonmail.com
Sources: Analysis based on official documentation from OpenAI, Google Cloud, and Inngest, along with publicly available announcements and technical specifications as of October 2025. Pricing and features subject to change—consult official documentation for current information.
Subscribe for Updates
Get the latest tools, insights, and updates from Agent Kits.