Claude Agent SDK+52 Strong positive

Operator Insights

What are teams learning from building on Claude Agent SDK and running it in production?

6 accounts: 1 failure, incident or rollback, 1 switched and 4 adopted it

Failures, incidents and rollbacks1

John Wall (juanwall)Builder, GitHub issue authorMoon Banking
Repetitive agent tasks run through the TypeScript SDK.
GitHub, anthropics/claude-agent-sdk-typescript, published Dec 6, 2025Read the write-up (opens GitHub, anthropics/claude-agent-sdk-typescript in a new tab)
Observation
Moon Banking's builder rolled back from SDK 0.1.60 to 0.1.30 when logged costs tripled; Anthropic replied that the older versions had undercounted client-side costs while billing stayed correct.
Takeaway
Budget from the Anthropic console or billing, not the SDK's client-side cost figures, which undercounted in older versions and sent this builder into a downgrade.
What happened
John Wall ran the same repetitive tasks before and after upgrading @anthropic-ai/claude-agent-sdk from 0.1.30 to 0.1.60 and reported that LLM cost tripled, consistently; downgrading to 0.1.30 cut the reported cost by two thirds (GitHub issue, Dec 6, 2025). Anthropic's reply of Dec 15, 2025 closed the issue: older versions undercounted client-side costs because of a bug, since fixed, and the Anthropic console and billing always gave the correct value.
- Public practitioner write-up (GitHub issue)
- Vendor reply on the issue
Deployment details
Repetitive agent tasks run through the TypeScript SDK.
- Public practitioner write-up (GitHub issue)
- Vendor reply on the issue
Team
Moon Banking
Workload
The same repetitive tasks before and after the upgrade
Product version
@anthropic-ai/claude-agent-sdk 0.1.60, rolled back to 0.1.30
Measurement
Logged cost tripled on 0.1.60; Anthropic says 0.1.30 undercounted it and billing was always correct
In SignalsCounted once, as the author's own issue. In the record

Switched to or from it1

SerafimKorablevAgent-infrastructure builder, Hacker News commenter21st.dev
21st.dev Agents, a multi-tenant web backend that runs agents for many users.
Hacker News, published Apr 7, 2026Read the write-up (opens Hacker News in a new tab)
Observation
21st.dev migrated its multi-tenant web agents from the Vercel AI SDK to the Claude Code SDK, then built sandboxing, streaming, threads and approval screens around it.
Takeaway
Budget the web platform around the SDK before adopting it for many tenants. Sandboxes, credential isolation, streaming, threads and approval screens are yours to build.
What happened
21st.dev started on the Vercel AI SDK and migrated to the Claude Code SDK for its full agent loop, then found the SDK built for a single-user CLI (Hacker News comment, Apr 7, 2026). The team built E2B sandboxing per session, environment isolation so credentials cannot leave the sandbox, SSE streaming, persistent threads, cost tracking, logs and session replay around it. Plan mode, ask-user and approval prompts exist only as CLI tools, so it rebuilt them as React components.
“…we went through this exact thing. started on vercel ai sdk, migrated to claude code sdk hoping to get the full agent loop out of the box. same conclusion - it's built for single-user cli, not multi-tenant web backends.”
SerafimKorablev
- Public practitioner write-up (Hacker News comment)
- No team metrics reported
Deployment details
21st.dev Agents, a multi-tenant web backend that runs agents for many users.
- Public practitioner write-up (Hacker News comment)
- No team metrics reported
Team
21st.dev engineering
Workload
Agents for many web users, each session in its own sandbox
Product version
Claude Code SDK, after the Vercel AI SDK
In SignalsCounted once, as the commenter's own account. In the record

Adopted it4

Niklas GustavssonChief Architect and VP of EngineeringSpotify
Fleet Management at Spotify. Migrations and maintenance across thousands of components, and tasks engineers hand to Honk in Slack.
Spotify Engineering, published Jun 3, 2026Read the write-up (opens Spotify Engineering in a new tab)
Observation
Spotify runs Honk, its background coding agent, on the Agent SDK inside its own harness, with many sessions scheduled in Kubernetes pods.
Takeaway
Wrap the SDK in your own harness and scheduler, and give the agent your CI as a tool so every change it makes is built and checked before a person sees it.
What happened
Deterministic migration scripts broke on complex code changes, so Spotify built Honk, a background coding agent. Honk runs Claude through the Agent SDK inside Spotify's own harness, deployed in Kubernetes pods so many sessions run at once. Its trusted tools include builds in Spotify's CI on several operating systems. Fleetshift picks targets and tracks progress while Honk makes the code changes. Spotify reports its latest Java migration across backend services took three days.
“Under the hood, Honk runs Claude using the Agent SDK, wrapped inside our own harness and deployed in Kubernetes pods so we can schedule many sessions concurrently across our cloud environment.”
Spotify Engineering
- Public engineering write-up (summary of a Code with Claude 2026 talk)
- Team metric reported
Deployment details
Fleet Management at Spotify. Migrations and maintenance across thousands of components, and tasks engineers hand to Honk in Slack.
- Public engineering write-up (summary of a Code with Claude 2026 talk)
- Team metric reported
Team
Spotify engineering, through Fleet Management and Honk in Slack
Workload
Fleet-wide code migrations and maintenance pull requests, plus tasks handed off in Slack
Product version
Claude through the Agent SDK inside Spotify's own harness, in Kubernetes pods
Measurement
The latest Java migration across backend services took three days
In SignalsCounted once, as Spotify Engineering's own post. In the record

CrystallizeOct 13, 2026Crystallize blogCrystallize's internal agent runs one SDK session per Slack request and keeps every credential with the orchestrator, outside the model's reach.
Sébastien MorelCTOCrystallize
Claudette, an internal Slack agent for Crystallize engineers. It answers code questions, drafts plans and pull requests, and checks documentation against the code.
Crystallize blog, published Oct 13, 2026Read the write-up (opens Crystallize blog in a new tab)
Observation
Crystallize's internal agent runs one SDK session per Slack request and keeps every credential with the orchestrator, outside the model's reach.
Takeaway
Treat the SDK session as untrusted: keep credentials in the orchestrator, enforce read-only access at each provider, and sandbox every shell command at the operating system.
What happened
Crystallize built an orchestrator that receives Slack requests, checks who is asking, and launches a Claude Code session through the Claude Agent SDK for each one. Investigations and code changes go to the strongest model and routine readings to a lighter one. Shell commands run in an operating-system sandbox, Seatbelt on macOS and bubblewrap on Linux, and file tools get the same list of forbidden paths. Credentials stay with the orchestrator, which runs privileged calls itself. An early allowlist of safe commands was dropped because one shell line can hide a second command. Crystallize's own attacks on an earlier version found that a git hook planted in the model's copy could run on the host during a push, and that changing the copy's remote could redirect the push; both worked until it closed them. Code arrives only as draft pull requests an engineer must merge.
“In an earlier version, a git hook planted in the model's copy could run on the host during the push, and changing the remote in that copy could redirect the push. Both attacks worked.”
Crystallize engineering
Public engineering write-up (CTO's account)
No team metrics reported
Deployment details
Claudette, an internal Slack agent for Crystallize engineers. It answers code questions, drafts plans and pull requests, and checks documentation against the code.
Public engineering write-up (CTO's account)
No team metrics reported
Team
Crystallize engineers, internal use only
Workload
Code questions in Slack, support co-pilot, plans, draft pull requests, and documentation checks
Product version
Claude Code sessions through the Claude Agent SDK, launched per request by Crystallize's orchestrator on Bun
In SignalsCounted once, as the CTO's own post. In the record
PostHogJun 22, 2026PostHog blogPostHog runs the Claude Agent SDK in sandboxes to research promoted problem reports and to write fixes, which it reruns until CI is green.
Josh SnyderCo-author with Cleo Lant, PostHog blogPostHog
PostHog's pipeline that groups product signals into reports and turns actionable ones into pull requests.
PostHog blog, published Jun 22, 2026Read the write-up (opens PostHog blog in a new tab)
Observation
PostHog runs the Claude Agent SDK in sandboxes to research promoted problem reports and to write fixes, which it reruns until CI is green.
Takeaway
Filter for specific, actionable problems before handing them to the SDK, and keep sandbox snapshots so a fix can resume when CI or a reviewer pushes back.
What happened
When a group of signals crosses a weight threshold, PostHog starts a research agent on the SDK in a sandbox, with PostHog's MCP server, the codebase, Linear and Notion, and it returns a problem summary, a priority and a suggested reviewer (PostHog blog, Jun 22, 2026). For actionable problems PostHog clones the repo into a sandbox, runs the SDK to build a fix and pushes a PR. It snapshots each sandbox and rehydrates it when CI fails or a reviewer comments. Vague reports are filtered out first, because the agent will change something regardless.
“Hand a generic report like "onboarding is broken" to the Agent SDK or Claude Code, and it'll go and change something regardless.”
Josh Snyder and Cleo Lant
Public engineering write-up
No team metrics reported
Deployment details
PostHog's pipeline that groups product signals into reports and turns actionable ones into pull requests.
Public engineering write-up
No team metrics reported
Team
PostHog product engineering
Workload
Research on promoted problem reports, and fix pull requests
Product version
Claude Agent SDK in sandboxes with snapshots
In SignalsCounted once, as PostHog's own post. In the record
Thomson ReutersMar 10, 2026Thomson Reuters Innovation blogThomson Reuters rebuilt CoCounsel Legal from the ground up around the Claude Agent SDK; the wider CoCounsel family has more than one million professional users.
Rawia AshrafHead of Product, CoCounsel Transactional, co-author with Emily Colbert, Head of Product, CoCounsel LitigationThomson Reuters
CoCounsel Legal, Thomson Reuters' AI assistant for lawyers, with native access to Westlaw and Practical Law.
Thomson Reuters Innovation blog, published Mar 10, 2026Read the write-up (opens Thomson Reuters Innovation blog in a new tab)
Observation
Thomson Reuters rebuilt CoCounsel Legal from the ground up around the Claude Agent SDK; the wider CoCounsel family has more than one million professional users.
Takeaway
The SDK can sit at the model layer of a regulated product. Thomson Reuters keeps its content, editors and governance around it, so the harness is one layer of the system.
What happened
Thomson Reuters re-architected CoCounsel Legal from the ground up to put Anthropic's Claude Agent SDK at the model layer (Thomson Reuters innovation blog, Mar 10, 2026). The company says more than one million professionals in 107 countries and territories use CoCounsel, and the new version was entering a US beta with general availability planned for later in 2026. Its CTO, Joel Hron, credits Claude with the reasoning and Thomson Reuters' own content, domain experts and governance with the outcomes customers get. The company describes its AI architecture as model-agnostic.
Public organization write-up (Thomson Reuters, an Anthropic partner)
No team metrics reported
Deployment details
CoCounsel Legal, Thomson Reuters' AI assistant for lawyers, with native access to Westlaw and Practical Law.
Public organization write-up (Thomson Reuters, an Anthropic partner)
No team metrics reported
Team
Thomson Reuters CoCounsel product teams
Duration
US beta from spring 2026, general availability planned for later in 2026
Workload
Legal research and drafting for lawyers on Westlaw and Practical Law content
Product version
Claude Agent SDK at CoCounsel Legal's model layer

Failures, incidents and rollbacks1

Moon Banking's builder rolled back from SDK 0.1.60 to 0.1.30 when logged costs tripled; Anthropic replied that the older versions had undercounted client-side costs while billing stayed correct.

Switched to or from it1

21st.dev migrated its multi-tenant web agents from the Vercel AI SDK to the Claude Code SDK, then built sandboxing, streaming, threads and approval screens around it.

Adopted it4

Spotify runs Honk, its background coding agent, on the Agent SDK inside its own harness, with many sessions scheduled in Kubernetes pods.

Crystallize's internal agent runs one SDK session per Slack request and keeps every credential with the orchestrator, outside the model's reach.

PostHog runs the Claude Agent SDK in sandboxes to research promoted problem reports and to write fixes, which it reruns until CI is green.

Thomson Reuters rebuilt CoCounsel Legal from the ground up around the Claude Agent SDK; the wider CoCounsel family has more than one million professional users.