WhatWeRun, a database of real production AI agent stacks

WhatWeRun launched an open catalog indexing real-world artificial intelligence agent architectures running in commercial environments, drawing technical specifications directly from public engineering posts, community disclosures, and company postmortems. According to whatwerun.dev's report, the public ledger documents 57 production stacks across software engineering, internal operations, customer support, and sales automation. Each catalog entry retains original operational constraints, preserving exact pricing baselines across per-task, per-run, or monthly accounting models without projecting estimated figures.

This report examines verified architectures, operational frameworks, and telemetry platforms operating across production engineering teams. Systems engineers, software architects, and platform managers can inspect how commercial teams structure autonomy boundaries, coordinate multi-agent workflows, and isolate code execution environments. The analysis details production telemetry, unit economics, infrastructure runtimes, and orchestration frameworks across modern commercial deployments.

01Verified agent architectures from engineering writeups

Production engineering teams extract architectures directly from primary disclosures, cataloging infrastructure components without modifying source data. In financial technology, Bkper operates an in-house gateway and command-line harness alongside Gemini Flash parsers, routing transactions through account discovery and document extraction while providing user environments inside ephemeral Linux sandboxes. Similarly, knowledge workers deploy long-running document processors on LangGraph, enabling agents to parse complex document batches for thirty minutes against expert operational playbooks before concluding runs autonomously.

Commercial healthcare and consumer data platforms configure specialized agents for internal monitoring and analytical discovery. Natera deploys SWARM lab-operations agents on Amazon Bedrock AgentCore Runtime to diagnose cfDNA testing support tickets, generating structured diagnostic reports on Jira issues before human engineers intervene. Meanwhile, Pinterest operates an internal Analytics Agent querying database tables, dashboards, and metrics, pairing queries with automated propose-test-learn loop engineering to reinforce procedural step compliance.

Corezoid utilizes ticket-routing coding agents managed through a digital twin from Simulator.Company, invoking a LiteLLM gateway to resolve labeled Jira issues and submit merge requests at a stated run cost of approximately 2.50 dollars per change. Amazon deploys three autonomous agent teams that author, test, and repair the downloadable AWS Emulator without human-authored or human-reviewed pull requests, scaling beyond initial Kiro prototypes. These deployments establish reproducible patterns for ticket triage and isolated software synthesis.

02Enterprise deployments and automated engineering workflows

Internal enterprise platforms expand beyond individual task assistants into dedicated multi-service environments. Workday launched an internal Agent Platform built on LangGraph and LangChain Agent Server, delivering authentication, security governance, and compliance auditing across human resources, finance, and corporate IT operations. Across enterprise infrastructure, NVIDIA engineered Agent Substrate, a zero-trust Kubernetes execution layer providing high-density sandboxes for coding assistants, reinforcement learning post-training environments, and short-lived GPU runtimes, donating the technology to the Cloud Native Computing Foundation Sandbox.

Go-to-market teams apply orchestrated agents directly to sales pipelines and customer engagement channels. LangChain operates an internal GTM Agent powered by Deep Agents, generating account plans, proof-of-value dossiers, proposal slide decks, and onboarding guides before prospect discussions. Temporal utilizes Kepler, an internal prospecting agent running Claude models to evaluate buying signals, match prospective leads, and generate targeted outreach messaging for business development representatives.

Desktop software and consumer platforms implement specialized agent loops across user files and application triage workflows. Dropbox deployed file-acting agents via Codex protocols and Pi harnesses to reorganize directory trees, rename assets, and collate shared file packages. For mobile applications, Duolingo operates the AI STR Fix bug agent on Claude Agent SDK with Claude Sonnet, investigating Shake-to-Report tickets by aggregating code history, tracking events, and reproduction steps for 0.50 to 6.80 dollars per bug ticket.

03Unattended coding minions and research systems

High-volume autonomous code generation operates at production scale inside payment infrastructure providers. Stripe deploys Stripe Minions using Goose, an internal fork of Block's goose engine combined with custom Blueprint state machines. Minions autonomously complete and merge over 1,300 pull requests weekly without human-written code, requiring human review before deployment at estimated operating expenses ranging from 3 dollars to 18 dollars per merged pull request.

Data engineering organizations apply thin orchestration wrappers over multi-repository systems to eliminate repetitive migration bottlenecks. Meta Engineering built custom thin orchestrators driven by workflow templates to automate data column onboarding across recommendation-system pipelines, coordinating six execution stages across three distinct code repositories and three programming languages. This architecture maintains deterministic control across heterogeneous production pipelines while minimizing orchestration overhead.

Frontier research labs and web-scale consumer platforms maintain dedicated agent architectures for unstructured synthesis. Anthropic deployed a multi-agent research architecture using custom orchestrator-worker layouts led by Claude Opus 4, querying web sources and Google Workspace integrations at estimated operational costs of 1.94 to 5.83 dollars per query. In independent deployments, ODSEA pairs LangGraph with Next.js 15 and Supabase to run continuous multi-user agent workflows twenty-four hours a day.

04Framework rankings across fifty-seven published setups

Aggregated metrics across all 57 indexed setups reveal distinct consolidation among open orchestration frameworks and cloud services. LangGraph leads framework adoption with 12 documented deployments, followed by Temporal with 6 implementations, Claude Agent SDK with 3 systems, and Deep Agents, Goose, and OpenAI Agents SDK recording 2 deployments each. Proprietary runtime layers such as Workday LangChain Python SDK, Agent Operating Procedures, and Agent Substrate appear in specialized operational environments.

Model utilization across indexed stacks exhibits balanced distribution among frontier foundation models. Claude and Claude Opus lead with 3 deployments each, tied alongside GPT-6 Astra at 3 implementations. Claude Haiku, Claude Sonnet, Claude Sonnet 4, GPT-5.5, and Jev each account for 2 production systems, supplemented by fallback providers routed through LiteLLM and specialized models simulating conversational patient evaluations.

Telemetry layers reflect diverse runtime monitoring strategies across distributed agent execution paths. LangSmith represents the most adopted observability tool with 5 deployments, while Datadog tracks 4 systems and Temporal monitors 3 architectures. OpenTelemetry accounts for 2 implementations, operating alongside enterprise solutions including Amazon CloudWatch, AWS X-Ray, Braintrust, Clearstone custom SQLite tracers, and proprietary AI Command Center usage dashboards.

05Cost transparency and weekly Tuesday digests

WhatWeRun maintains strict provenance standards across cataloged agent architectures, forbidding speculative cost projections when source engineering teams omit infrastructure expenditures. When teams publish financial metrics, entries retain the author's accounting framework, distinguishing per-task operational costs, per-run executions, and broader monthly infrastructural allocations. Older technical writeups carry explicit chronological identifiers, preventing engineers from misinterpreting legacy prototypes as modern production systems.

To maintain continuous platform tracking, the registry curates a weekly Tuesday email digest delivering five newly documented stacks directly to subscribers. Each digest summarizes architectural choices, links directly to source engineering writeups, and analyzes newly observed operational tooling patterns without publishing speculative industry commentary. This ongoing monitoring establishes empirical benchmarks for engineering organizations deploying production agents across cloud infrastructure.

The submission catalog remains open to community developers, infrastructure engineers, and research teams publishing verified agent implementations. Submissions require verifiable documentation from company blogs, public repositories, or first-person engineering accounts detailing real production constraints. By documenting actual model choices, orchestrators, and execution boundaries, the catalog provides a factual baseline for real enterprise agent performance.