Amazon Bedrock+9 Mixed

Operator Insights

What are teams learning from building on Amazon Bedrock and running it in production?

13 accounts: 2 failures, incidents or rollbacks, 1 switched, 3 run it beside another tool and 7 adopted it

Failures, incidents and rollbacks2

Ziva engineeringCompany engineering blogZiva
Small SaaS selling AI tools for the Godot game engine. Claude Haiku, Sonnet and Opus on Bedrock in us-west-2 through cross-Region inference profiles.
Ziva blog, published May 16, 2026Read the write-up (opens Ziva blog in a new tab)
Observation
At Ziva, Bedrock quotas 1,000 to 2,000 times below AWS's defaults throttled 14% to 16% of Claude requests, and 16 days of escalation failed to raise them.
Takeaway
Read the quotas your account was actually given before you build on Bedrock, and keep a second route to the same models ready, since a quota request can fail regardless of support tier.
What happened
Ziva's account was created with quotas 1,000 to 2,000 times below AWS's documented defaults: 1,000 times below for Haiku and Sonnet, and 2,000 times below for Opus 4.6, at 5 requests per minute against a 10,000 default. CloudWatch showed 14% of Haiku requests and 16% of Opus requests throttled over 14 days. The team escalated through four channels, including paid Business Support+, and after 16 days AWS still had not raised the quota (May 16, 2026).
“CloudWatch’s AWS/Bedrock namespace showed a 14% throttle rate on Haiku and 16% on Opus over the 14 days preceding the case.”
Ziva.sh
- Public engineering write-up
- Throttle rates read from CloudWatch over 14 days
Deployment details
Small SaaS selling AI tools for the Godot game engine. Claude Haiku, Sonnet and Opus on Bedrock in us-west-2 through cross-Region inference profiles.
- Public engineering write-up
- Throttle rates read from CloudWatch over 14 days
Team
Ziva engineering
Duration
16 days of quota escalation
Workload
Small SaaS workload in us-west-2
Product version
Bedrock cross-Region inference profiles
Model
Claude Haiku 4.5, Sonnet 4.5 and Opus 4.6
Measurement
14% of Haiku and 16% of Opus requests throttled over 14 days (May 2026)

PositJul 27, 2026GitHub (posit-dev/ai-lib)At Posit, every message to Claude Fable 5 on Bedrock failed with HTTP 400 because the default data retention mode was not available for the model, and the cause was still unconfirmed in October.
sharon-wangPosit engineer by her GitHub profile, contributor to posit-dev/ai-libPosit
Posit Assistant 0.8.0 in the Positron IDE, with Amazon Bedrock as one model provider.
GitHub (posit-dev/ai-lib), published Jul 27, 2026Read the report (opens GitHub (posit-dev/ai-lib) in a new tab)
Observation
At Posit, every message to Claude Fable 5 on Bedrock failed with HTTP 400 because the default data retention mode was not available for the model, and the cause was still unconfirmed in October.
Takeaway
Send a test message to each model under your account's retention mode in every Region before listing it to users: Posit's users met a raw HTTP 400 on every Fable 5 message, and three months later its engineers had still not confirmed the cause.
What happened
Selecting Claude Fable 5 on Posit Assistant's Bedrock provider returned HTTP 400 on every message before generation, with an error saying the default data retention mode was not available for the model (GitHub issue, Jul 27, 2026). Posit's Bedrock requests set no retention parameter, so the default came from the account or the SDK. A later comment proposed keeping such models out of the picker. On October 23, 2026 the issue was still open: whether the account's retention setting causes it, and which retention modes Fable 5 accepts on Bedrock, were unconfirmed.
“Selecting Claude Fable 5 on the Amazon Bedrock provider and sending any message returns HTTP 400 before generation…”
sharon-wang
Public engineering write-up (GitHub issue in Posit's own repository)
Failure reproduced on every request
Deployment details
Posit Assistant 0.8.0 in the Positron IDE, with Amazon Bedrock as one model provider.
Public engineering write-up (GitHub issue in Posit's own repository)
Failure reproduced on every request
Team
Posit Assistant engineers
Duration
Open from Jul 27 to at least Oct 23, 2026
Workload
AI assistant in the Positron IDE
Product version
Posit Assistant 0.8.0, Amazon Bedrock provider
Model
Claude Fable 5
Measurement
Every message returned HTTP 400

Switched to or from it1

Shaurya KethireddyLead Member of Technical StaffSlack
Slack AI serving infrastructure, which serves millions of users.
Engineering at Slack, published May 28, 2026Read the write-up (opens Engineering at Slack in a new tab)
Observation
Slack moved Slack AI from its own SageMaker GPU endpoints to Bedrock, and in 2026 added Vertex AI as a second cloud.
Takeaway
Split Bedrock capacity by workload, reserved for interactive features and on demand for batch jobs, and keep a second cloud behind the same serving layer so one provider cannot take the feature down.
What happened
From mid-2024 Slack migrated Slack AI's model serving from self-managed SageMaker endpoints to Bedrock, using Provisioned Throughput for latency-sensitive features such as channel summaries and On Demand for bursty jobs such as Recap. Feature flags shifted traffic gradually with instant rollback, and Slack reports 100% availability during the switch. New models reached Bedrock weeks or months before SageMaker. By early 2026 Slack added Google Cloud Vertex AI beside Bedrock for redundancy and a broader catalog, reporting about 67% lower latency on high-velocity, low-token workloads from matching models to features (May 28, 2026).
“We utilized PT for interactive, latency-sensitive features like channel summaries, while leveraging OD for bursty, scheduled workloads like Recap to eliminate costs for idle compute.”
Shaurya Kethireddy
- Public engineering write-up
- Availability and latency figures self-reported
Deployment details
Slack AI serving infrastructure, which serves millions of users.
- Public engineering write-up
- Availability and latency figures self-reported
Team
Slack AI infrastructure
Duration
Since mid-2024; Vertex AI added by early 2026
Workload
Slack AI features such as channel summaries and Recap
Product version
Bedrock Provisioned Throughput and On Demand, with Vertex AI as a second cloud
Measurement
100% availability during the switch; about 67% lower latency on low-token workloads (May 2026)

Run it beside another tool3

Siddharth KodwaniSoftware Engineer, GenAI PlatformDoorDash
DoorDash's GenAI platform team, serving product teams across the company through one LLM Gateway.
InfoQ, published Oct 3, 2026Watch the talk (opens InfoQ in a new tab)
Observation
DoorDash runs the Claude API with Amazon Bedrock as its fallback, behind one gateway that attributes usage costs to each department.
Takeaway
Treat Bedrock as a second route to the same Claude models, wired in before an outage or a quota cut forces it, with cost attribution in the gateway from the first team onward.
What happened
DoorDash put OpenAI, Anthropic and Gemini models behind one API and SDK so product teams could try a new model with one change. For reliability each provider got a second route: the Claude API endpoint falls back to Amazon Bedrock, and OpenAI to Azure (InfoQ, Oct 3, 2026). As usage grew, finance asked who was spending what on shared keys, so the gateway added workspaces with cost attribution per department.
“We actually had to go understand what the fallback should be, what kind of routing should be. At one point, I remember we had OpenAI endpoint, but we also had a fallback on Azure endpoint. Or, we had Claude model API endpoint, then we had a fallback on Bedrock, so that the services can rely on us from the reliability point of view. We went in that direction. We quickly realized that as the usage is going up, then we need a way to track, because now finance got involved. Like, you guys have one API key for multiple teams, how do I track who is using what? We quickly realized we need workspaces where we have cost attribution. We can actually drill down on who is using how much. We can attribute cost back to those departments. That's what basically we did with LLM Gateway.”
Siddharth Kodwani
- Conference talk (InfoQ)
- No metrics reported
Deployment details
DoorDash's GenAI platform team, serving product teams across the company through one LLM Gateway.
- Conference talk (InfoQ)
- No metrics reported
Team
DoorDash GenAI Platform
Workload
Hundreds of product use cases behind one LLM Gateway
Product version
Amazon Bedrock as the fallback for the Claude API endpoint
Model
OpenAI, Anthropic and Gemini models

IntercomAug 8, 2025Fin AI ResearchFin serves Anthropic models through Bedrock, Vertex and Anthropic's own API, and moved traffic between providers during an OpenAI outage.
Fin AI research teamResearch teamIntercom
Fin, Intercom's customer service agent, built on third-party models behind its own reliability layer.
Fin AI Research, published Aug 8, 2025Read the write-up (opens Fin AI Research in a new tab)
Observation
Fin serves Anthropic models through Bedrock, Vertex and Anthropic's own API, and moved traffic between providers during an OpenAI outage.
Takeaway
Use Bedrock as one of several hosts for the same model family, so an outage moves traffic and keeps the product running.
What happened
Fin routes each model across more than one host: Anthropic models through AWS Bedrock, Google Vertex or Anthropic's own infrastructure, and OpenAI models through Azure or OpenAI. During the OpenAI outage of June 10, 2025 it moved traffic from OpenAI to Anthropic models (Aug 8, 2025).
“For example, Anthropic models can be served by AWS Bedrock, GCP Vertex, or Anthropic’s own infrastructure; OpenAI models are served by either Azure or OpenAI itself.”
Fin AI research team
Public organization write-up
No metrics reported
Deployment details
Fin, Intercom's customer service agent, built on third-party models behind its own reliability layer.
Public organization write-up
No metrics reported
Team
Fin AI
Workload
Customer service agent serving Intercom customers
Product version
Anthropic models on Bedrock beside Vertex and Anthropic's API
Model
Anthropic and OpenAI models
WealthsimpleFeb 6, 2025InfoQWealthsimple added Bedrock to its internal LLM gateway beside OpenAI, Cohere and Gemini, then rebuilt the gateway on OpenAI's API format.
Mandy GuAuthor of InfoQ's article on Wealthsimple's generative AI toolsWealthsimple
Wealthsimple's internal LLM gateway for employee productivity tools.
InfoQ, published Feb 6, 2025Read the write-up (opens InfoQ in a new tab)
Observation
Wealthsimple added Bedrock to its internal LLM gateway beside OpenAI, Cohere and Gemini, then rebuilt the gateway on OpenAI's API format.
Takeaway
Add Bedrock behind a gateway that speaks one API format, so each new provider is a configuration change for the teams that call it.
What happened
Wealthsimple's gateway started with OpenAI and Cohere, added Gemini for its 1M-token context, then added Amazon Bedrock. Integrating each provider's interface caused friction, so in September 2024 the team shipped version 2 of its API mirroring OpenAI's specification (InfoQ, Feb 6, 2025).
“After supporting multi-modal inputs, the next thing we added to our platform was Amazon Bedrock. Bedrock is AWS's managed service for interacting with foundational large language models.It also provides the ability to deploy and fine-tune these models at scale.”
Mandy Gu
Public engineering write-up (InfoQ article by Wealthsimple's engineering leader)
No metrics reported
Deployment details
Wealthsimple's internal LLM gateway for employee productivity tools.
Public engineering write-up (InfoQ article by Wealthsimple's engineering leader)
No metrics reported
Team
Wealthsimple machine learning and platform engineering
Duration
2023 to September 2024
Workload
Internal gateway for employee productivity tools
Product version
Amazon Bedrock beside OpenAI, Cohere and Gemini, behind an OpenAI-format API

Adopted it7

NeradotAug 31, 2026Neradot blogAt Neradot, a support agent's Bedrock bill fell 36% once the stable part of the prompt moved to the top and was cached.
Liza KatzAuthor of Neradot's write-upNeradot
Agency engineering team. ReAct support agent for a gaming company's VIP players. In production.
Neradot blog, published Aug 31, 2026Read the write-up (opens Neradot blog in a new tab)
Observation
At Neradot, a support agent's Bedrock bill fell 36% once the stable part of the prompt moved to the top and was cached.
Takeaway
Shrink the prompt before caching it and put stable content first. Count reads per cache write before adding a breakpoint. Reuse across requests pays. Reuse within one short ticket barely does.
What happened
Neradot built a human-reviewed ReAct agent for a gaming company's VIP support on Claude through Bedrock. Each loop resent the full prompt. A fixed block of persona, procedures, and tool schemas used about 14,000 tokens and made up 76% of each ticket's first request. The agent resent it a few hundred thousand times a day. Compaction cut prompt tokens 15%, then one cache breakpoint covered the fixed block. The team removed a second breakpoint after ticket context because each ticket made only 2 or 3 requests. The two passes cut the bill 36% with unchanged outputs and evaluation scores.
“Two passes later the bill was down 36%, with no measurable change in output quality.”
Liza Katz
Public engineering write-up
Cost reduction and token shares measured
Model family stated
Deployment details
Agency engineering team. ReAct support agent for a gaming company's VIP players. In production.
Public engineering write-up
Cost reduction and token shares measured
Model family stated
Team
Neradot engineering, building for a gaming company client
Workload
VIP customer support agent with human review
Product version
Claude on Amazon Bedrock with one explicit cache breakpoint
Model
Claude
Measurement
Inference cost down 36%; prompt tokens down 15% from compaction
ZalandoSep 25, 2025Zalando Engineering BlogZalando's SRE team runs postmortem analysis on Claude Sonnet 4 in Bedrock, chosen for compliance, and one finding now blocks a quarter of later datastore incidents.
Dmitry KolesnikovSenior Principal EngineerZalando
Zalando SRE, an LLM pipeline that reads thousands of datastore incident postmortems.
Zalando Engineering Blog, published Sep 25, 2025Read the write-up (opens Zalando Engineering Blog in a new tab)
Observation
Zalando's SRE team runs postmortem analysis on Claude Sonnet 4 in Bedrock, chosen for compliance, and one finding now blocks a quarter of later datastore incidents.
Takeaway
When compliance blocks a vendor API, Bedrock can carry the workload inside the existing AWS approval; judge the pipeline by the incidents its findings prevent.
What happened
Zalando prototyped on small local models, which hallucinated up to 40% of the time, then moved the pipeline to Claude Sonnet 4 on Bedrock, mainly for compliance once legal approved cloud-hosted models for data with personal information. One pattern it surfaced led to automated validation of infrastructure-as-code changes, which Zalando says shields it from 25% of subsequent datastore incidents (Sep 25, 2025).
“This insight directly led to the solution for automated change validation for infrastructure as code which is able to shield us from 25% subsequent datastore incidents…”
Dmitry Kolesnikov
Public engineering write-up
Incident reduction self-reported
Deployment details
Zalando SRE, an LLM pipeline that reads thousands of datastore incident postmortems.
Public engineering write-up
Incident reduction self-reported
Team
Zalando Site Reliability Engineering
Duration
Two years of postmortems analyzed
Workload
Analysis of thousands of datastore postmortems
Product version
Claude on Amazon Bedrock, after local open models
Model
Claude Sonnet 4
Measurement
25% of subsequent datastore incidents shielded by the resulting change validation (Sep 2025)
QlikOct 7, 2026AWS Machine Learning Blog (vendor case study)At Qlik, every Bedrock request and answer passes through Guardrails, and a model missing in a Region runs elsewhere until it arrives.
Kent McLeanSenior Director of EngineeringQlik
Product engineering. Qlik Answers. More than 40,000 customers (as of Oct 7, 2026) across 11 AWS Regions.
AWS Machine Learning Blog (vendor case study), published Oct 7, 2026Read the case study (opens AWS Machine Learning Blog (vendor case study) in a new tab)
Observation
At Qlik, every Bedrock request and answer passes through Guardrails, and a model missing in a Region runs elsewhere until it arrives.
Takeaway
Put a gateway in front of Bedrock that verifies each answer against the documents it cites. Plan capacity per Region months ahead, with a fallback for models that arrive late.
What happened
Qlik Answers reaches Bedrock through Qlik's LLM gateway for chat, streaming, embeddings, and reranking. Guardrails filter requests and responses for prompt injection, personal data, secrets, and denied topics, then check answers against their sources. Cross-Region inference serves 11 Regions under residency rules. Qlik uses SageMaker AI when a required model has not reached a Region, then moves back to Bedrock. The team plans token use and model availability 3 to 6 months before major launches.
Vendor case study
Written with Qlik's Senior Director of Engineering on AWS's Machine Learning Blog
Customer outcomes reported
Deployment details
Product engineering. Qlik Answers. More than 40,000 customers (as of Oct 7, 2026) across 11 AWS Regions.
Vendor case study
Written with Qlik's Senior Director of Engineering on AWS's Machine Learning Blog
Customer outcomes reported
Team
Qlik engineering
Duration
Generally available since February 2026
Workload
Grounded answers over structured and unstructured enterprise data
Product version
Bedrock through Qlik's LLM gateway, with Guardrails and cross-Region inference
Measurement
More than 100,000 discoveries from Qlik's Discovery Agent since launch (as of Oct 7, 2026)
SlackJan 14, 2026AWS Big Data Blog (vendor case study)At Slack, Bedrock supplies the models for Spark tuning advice so job metrics never leave Slack's AWS account.
Nilanjana MukherjeeStaff Software EngineerSlack
Data platform engineering. Apache Spark on Amazon EMR. Terabytes of data a day.
AWS Big Data Blog (vendor case study), published Jan 14, 2026Read the case study (opens AWS Big Data Blog (vendor case study) in a new tab)
Observation
At Slack, Bedrock supplies the models for Spark tuning advice so job metrics never leave Slack's AWS account.
Takeaway
When the data behind an AI feature is sensitive, keep model calls inside the cloud account that already holds the data. Put the effort into the metrics and the prompt structure the model reads.
What happened
Slack's framework captures more than 40 metrics across Spark applications, jobs, stages, and tasks. An internal MCP server exposes them with a structured tuning prompt to developers in Cursor or Claude Code. Because the metrics can be sensitive, Slack uses Bedrock models inside its own AWS account. Slack reports compute cost up to 50% lower and jobs more than 40% faster. Developer time spent tuning fell by more than 90%.
Vendor case study
Written with Slack engineers on AWS's Big Data Blog
Results reported as ranges
Deployment details
Data platform engineering. Apache Spark on Amazon EMR. Terabytes of data a day.
Vendor case study
Written with Slack engineers on AWS's Big Data Blog
Results reported as ranges
Team
Slack data platform and data foundations engineers
Workload
AI tuning recommendations for Apache Spark jobs on Amazon EMR
Product version
Amazon Bedrock in Slack's AWS account, behind an internal MCP server
Measurement
Compute cost up to 50% lower, jobs over 40% faster, tuning time down over 90%
HoneycombJul 6, 2026Honeycomb blogHoneycomb runs its Canvas investigation agents on Bedrock AgentCore, and lost session history on every deploy until it moved state to S3 Files.
Moses MendozaAuthor, Honeycomb engineering blog; co-written with Staff Software Engineer Martin HolmanHoneycomb
Honeycomb Canvas, a collaborative incident investigation agent.
Honeycomb blog, published Jul 6, 2026Read the write-up (opens Honeycomb blog in a new tab)
Observation
Honeycomb runs its Canvas investigation agents on Bedrock AgentCore, and lost session history on every deploy until it moved state to S3 Files.
Takeaway
On AgentCore, keep agent state outside the runtime from the first deploy; sessions do not survive a version update.
What happened
Honeycomb runs Canvas on AgentCore Runtime and Agent Memory. A new Runtime version sends live sessions to fresh environments, and session storage does not survive version updates, so Canvas lost its in-memory coordination history on every deploy. The team pinned each investigation to an endpoint and moved session history to S3 Files (Jul 6, 2026).
“Since Canvas's active coordination state lives in memory, we were losing history on every deploy.”
Moses Mendoza
Public engineering write-up
No metrics reported
Deployment details
Honeycomb Canvas, a collaborative incident investigation agent.
Public engineering write-up
No metrics reported
Team
Honeycomb Canvas engineering
Workload
Incident investigation agents
Product version
AgentCore Runtime and Agent Memory, session history in S3 Files
TinesMar 10, 2026Tines blogTines built its first AI features on Bedrock to keep customer data in each tenant's Region and out of model training.
Yannick GlosterSoftware EngineerTines
Tines workflow automation platform, AI features served to customer tenants.
Tines blog, published Mar 10, 2026Read the write-up (opens Tines blog in a new tab)
Observation
Tines built its first AI features on Bedrock to keep customer data in each tenant's Region and out of model training.
Takeaway
Start on Bedrock when tenant residency decides the first launch, and design the provider layer so a second wire format needs only an adapter.
What happened
Tines launched its first AI features on Bedrock so customer data stayed in the same geographic Region as each tenant and out of any training pipeline. When it later added Anthropic and OpenAI-compatible providers, Bedrock's Converse wire format and tool schema needed their own translation layer beside the others (Mar 10, 2026).
“The first AI features we launched were built on AWS Bedrock in order to always maintain customer data within the same geographic regions as their tenant, as well as out of any LLM training pipelines.”
Yannick Gloster
Public engineering write-up
No metrics reported
Deployment details
Tines workflow automation platform, AI features served to customer tenants.
Public engineering write-up
No metrics reported
Team
Tines engineering
Workload
AI features inside the Tines automation platform
Product version
Bedrock Converse API beside Anthropic and OpenAI-compatible clients
OpenSearchMay 7, 2026GitHub (opensearch-project)The OpenSearch project runs AI pull request review on Bedrock models across its main repository.
Mohit KumarMaintainer of opensearch-project/cross-cluster-replication, affiliation Amazon in its MAINTAINERS fileOpenSearch
The open source OpenSearch project's repositories on GitHub. Amazon started OpenSearch, and the request's author lists Amazon as employer, so this is an account from inside the vendor.
GitHub (opensearch-project), published May 7, 2026Read the report (opens GitHub (opensearch-project) in a new tab)
Observation
The OpenSearch project runs AI pull request review on Bedrock models across its main repository.
Takeaway
Bedrock can back a shared review workflow that every repository in a project reuses, billed to one AWS account.
What happened
OpenSearch's main repository runs a Code-Diff-Reviewer built on the open source PR-Agent with Amazon Bedrock models, shared as reusable workflows in opensearch-build. A May 7, 2026 request to extend it to cross-cluster-replication says it catches null pointer risks, wrong allocation IDs and missing error handling before human reviewers look.
“The main OpenSearch repo has this enabled and it catches issues like null pointer risks, incorrect allocation ID usage, missing error handling, and architectural concerns before human reviewers even look at the PR.”
mohit10011999
Public organization write-up (GitHub issue in the OpenSearch project, filed by an Amazon engineer)
No metrics reported
Deployment details
The open source OpenSearch project's repositories on GitHub. Amazon started OpenSearch, and the request's author lists Amazon as employer, so this is an account from inside the vendor.
Public organization write-up (GitHub issue in the OpenSearch project, filed by an Amazon engineer)
No metrics reported
Team
OpenSearch project maintainers
Workload
AI review of pull requests before human review
Product version
PR-Agent workflows in opensearch-build, powered by Amazon Bedrock

Failures, incidents and rollbacks2

At Ziva, Bedrock quotas 1,000 to 2,000 times below AWS's defaults throttled 14% to 16% of Claude requests, and 16 days of escalation failed to raise them.

At Posit, every message to Claude Fable 5 on Bedrock failed with HTTP 400 because the default data retention mode was not available for the model, and the cause was still unconfirmed in October.

Switched to or from it1

Slack moved Slack AI from its own SageMaker GPU endpoints to Bedrock, and in 2026 added Vertex AI as a second cloud.

Run it beside another tool3

DoorDash runs the Claude API with Amazon Bedrock as its fallback, behind one gateway that attributes usage costs to each department.

Fin serves Anthropic models through Bedrock, Vertex and Anthropic's own API, and moved traffic between providers during an OpenAI outage.

Wealthsimple added Bedrock to its internal LLM gateway beside OpenAI, Cohere and Gemini, then rebuilt the gateway on OpenAI's API format.

Adopted it7

At Neradot, a support agent's Bedrock bill fell 36% once the stable part of the prompt moved to the top and was cached.

Zalando's SRE team runs postmortem analysis on Claude Sonnet 4 in Bedrock, chosen for compliance, and one finding now blocks a quarter of later datastore incidents.

At Qlik, every Bedrock request and answer passes through Guardrails, and a model missing in a Region runs elsewhere until it arrives.

At Slack, Bedrock supplies the models for Spark tuning advice so job metrics never leave Slack's AWS account.

Honeycomb runs its Canvas investigation agents on Bedrock AgentCore, and lost session history on every deploy until it moved state to S3 Files.

Tines built its first AI features on Bedrock to keep customer data in each tenant's Region and out of model training.

The OpenSearch project runs AI pull request review on Bedrock models across its main repository.