AI agents rarely fail with an error. Their decisions get worse while dashboards stay green.
Mallika Rao's #InfoQ article recommends bounded autonomy, decision tracing and deterministic fallbacks for #AgenticAI.
A #WorkflowOrchestration layer enforces them around the agent.
An AI agent refunds a customer. Which system tells it the balance, and which one books the refund? #ApacheKafka and the database together are the source of truth for #AIAgents.
My new article:
kai-waehner.de/blog/2026/11/…
3
4
283
#Databricks bought Neon, #Snowflake bought Crunchy Data, and the clouds launched new #Postgres services for #AIAgents.
Why the bet landed on Postgres, and why #Oracle, SAP, IBM, and #OceanBase bet on their own engines instead:
kai-waehner.de/blog/2026/11/…
Stolen cloud and API credentials now cost victims over $100K in model charges in a single day. Mandiant describes how #LLMjacking works.
Budget caps, scoped API keys and alerts on newly enabled model services belong in every #AgenticAI architecture.
kai-waehner.de/blog/2026/11/…
117
#ApacheSpark 4.3 adds stateful processing to Real-Time Mode and moves closer to #ApacheFlink for low-latency streaming.
It fits when data and consumers live in the lakehouse.
Define the SLA first, then choose the engine:
kai-waehner.de/blog/2026/07/…
2
258
From October 2027, the amended EU261 rules give #Airlines 3 hours to reroute and 7 days to refund.
My new post maps one #WorkflowOrchestration control plane across data, infrastructure, apps and processes, with #ApacheKafka detecting the disruption:
kai-waehner.de/blog/2026/11/…
1
1
129
My #jaxcon keynote in front of around 600 people in Munich covered how enterprises keep control of #AgenticAI in production.
The harness controls one agent, and production needs end-to-end governance across data, infrastructure, apps and people. Thanks to the #Kestra booth team.
1
142
I get many #IndustrialIoT questions, so I consolidated my work on one page, from #OPCUA and #MQTT to #ApacheKafka at the IT/OT boundary, with the landscape, use cases and talks:
kai-waehner.de/industrial-io…
1
3
179
An approval rule in a system prompt works most of the time. Nobody can prove it worked every time.
Where retries, approvals, budgets and audit belong in an #AgenticAI architecture, and why #WorkflowOrchestration owns most of them:
kai-waehner.de/blog/2026/11/…
2
3
162
Apache Flink Agents 0.4 adds one API for TypeSafe’s Jev, OpenAI and Alibaba Cloud.
#DecisionModels now score and classify each event inside #ApacheFlink, with Trace Log lineage for every model and tool call.
kai-waehner.de/blog/2026/10/…
3
271
The #Mainframe still runs the core of many banks.
#DataStreaming its changes into #ApacheKafka with CDC brings that data into real time while the COBOL stays untouched.
OCC cut margin calculation from 8 hours to 5 minutes.
kai-waehner.de/blog/2025/06/…
7
343
Moving off Control-M and vRO/vRA was a big topic with financial services customers in New York this week, next to regulated workflows, deployment control and governed #AgenticAI.
All under one control plane with #Kestra.
Next US trip is #AWSreInvent in Vegas, let’s meet!
2
169
My new #Kestra whitepaper explains why infrastructure automation is not infrastructure orchestration. Terraform, Ansible and Argo CD each do their job, but none sequences the work across the others, and #PlatformEngineering teams feel that gap on Day 2.
kestra.io/resources/whitepap…
1
123
McKinsey expects humanoids to hold about 2% of #PhysicalAI value by 2045. The rest sits after the robot: quality holds, work orders, approvals.
New post on #DataStreaming and #WorkflowOrchestration for physical AI in manufacturing, logistics and energy:
kai-waehner.de/blog/2026/10/…
2
1
186
#ApacheKafka 4.4 is out.
Highlights:
1. #QueuesForKafka: Share groups get a dead-letter queue,
2. KRaft (ZooKeeper removal) bootstrapping gets simpler,
3. #KafkaStreams adds transactional state stores.
#OpenSource #DataStreaming
New on demand webinar with @mimacom .
Trusted AI starts with the data stream.
Why stale data now costs a wrong decision, how to govern events like APIs, and how @confluentinc fits next to IBM MQ, webMethods and watsonx.
#ApacheKafka #AgenticAI
mimacom.com/resources/truste…
1
1
138
Five weeks after #Jev:
AWS, Microsoft, Cloudflare, OpenAI, Snowflake and a dozen open source projects ship #DecisionModels on the same request schema.
The AI model is interchangeable. What it means for enterprise architecture:
kai-waehner.de/blog/2026/10/…
1
193
Great evening at the Java Vienna Meetup with event-driven architecture on #ApacheKafka, #UnifiedOrchestration and a live #Kestra demo.
Thanks to Dominik Dorn for organizing and for his functional #Java deep dive. Next time more Java, like the plugin SDK and Micronaut.
2
205
Flink CDC 3.7 puts LLM calls into the CDC pipeline.
Embeddings at ingestion make sense. Per-record LLM calls are paid again on every replay and return different answers, and LLM masking of PII will not pass an audit. Decision models like Jev fit the stream better.
#ApacheFlink
2
4
312
With MCP everywhere, calling an enterprise API is easy.
Access control is hard, because an agent acts without asking whether it is allowed to. Transaction identity belongs outside the agent.
#AgenticAI and #UnifiedOrchestration on Six Five Media:
sixfivemedia.com/content/why…
2
2
201




















