Key findings
- The harness earns the most consistent praise: all 21 sources that discuss Claude Code's tools, subagents, and permission controls in the SDK praise them.
- Thirteen sources report SDK agents in daily use, including Spotify's Honk in Kubernetes pods, Crystallize's Claudette, PostHog's fix-writing agent, a GitLab reviewer, and Slack agents.
- The CLI subprocess and Claude Code's defaults each draw concerns in 12 sources, and sessions on local disk in 10. One load test put P95 latency at 103 seconds at 40 concurrent requests.
- Prompt cache misses and cost increases appear in 9 sources, including measured figures.
- Plan billing draws concerns in 8 of 10 sources. Builders measured SDK runs using 1.7 to 1.8 times more of the 5-hour window than Claude Code in September and October 2026.
- Of 18 sources that describe a runtime choice, 14 kept the SDK and 4 wrote or adopted another loop.
The public accounts behind the Claude Agent SDK's Signals explain what builders get from the harness and the work it leaves them. The English-language sources span June 2025 to October 2026, from Thomson Reuters' CoCounsel rebuild, Spotify's background coding agent and Crystallize's internal Slack agent to GitHub issues, Reddit posts and solo developers.
What builders keep it for
Builders name the harness they did not have to write. Omnara's co-founder replaced terminal-output parsing with the SDK and calls it far easier to maintain. A Terminal Use co-founder who hosts agents from several SDKs singles out its hooks and permission protocol. Kyle Redelinghuys values subagents with their own context and sessions that resume and fork.
Production use spans engineering, internal and customer-facing work. Thomson Reuters re-architected CoCounsel Legal to put the SDK at its model layer, Spotify runs Honk inside its own harness in Kubernetes pods, and Crystallize starts one session per Slack request behind an orchestrator that holds the credentials. PostHog runs it in sandboxes to research problem reports and write fixes. Builders on Hacker News, Reddit and X describe a GitLab reviewer, a UAT agent driving Playwright, a WhatsApp assistant, and about ten Slack agents in production.
Where it adds work
The subprocess draws the most specific complaints. A team evaluating it for a Python web backend cites about 12 seconds of spawn time per call and memory growth, and a load test on a 4-core, 8 GB server put P95 latency at 103 seconds at 40 concurrent requests. Glide's VP of Engineering reports workarounds to run it in a worker runtime, and GitHub issues describe truncated pipes, a 950 MB session listing, and the wrong binary picked on Linux.
Sessions come next. Before Anthropic documented a SessionStore adapter, teams on Kubernetes synced session files to PostgreSQL by hand, and one saw resume silently start a new session. Builders of multi-tenant web products report rebuilding streaming, threads, and approval UI around the SDK.
Cost, caching, and plan terms
Cost reports center on caching. One builder measured cache reads on turn two rising from 10.5k to 50.7k of 50.8k tokens after passing a stable session ID. Another found the bundled CLI wrote cache on every turn without reading it, and a third saw cost triple between two minor versions.
Plan billing splits the record. Some builders run the SDK on a Claude plan at a predictable monthly cost; others could not tell whether their use was allowed, or faced separate usage billing. Anthropic's October 7, 2026 update says Max and Team plans include monthly API credits that cover the SDK, and SDK use can still draw on plan limits. Later that month builders measured SDK runs using 1.7 to 1.8 times more of the 5-hour window than Claude Code, and one found a hit quota returned as a successful result, so a job was marked complete.
Why some builders write their own loop
Four sources moved away from it or wrote a replacement. A web assistant team weighed the Messages API, an author found it too heavy to embed, and a builder needed a vendor-neutral harness. A testing company called the CLI binary directly after spawn bugs. Their reasons are weight, model lock-in, and release regressions, and none of them disputes what the harness does.
- Does the agent need files and a shell, or only your APIs?
- Can your runtime host one process per session?
- Will you need a model outside Claude?
- Who owns sessions, sandboxing, and approvals around it?
Methodology
GILD synthesis of the 100 English-language public accounts in the Claude Agent SDK's Signals, read and archived on October 25, 26 and 28, 2026 and published between June 2025 and October 2026. They comprise 27 Hacker News comments and posts, 33 GitHub issues, 32 of them on Anthropic's two SDK repositories, 15 posts on X, 5 Reddit posts, 15 company and personal blog posts, 3 podcast episodes, a conference workshop and a video. Each record keeps the exact words it was coded from. The 10 records from GILD's targeted search for problem reports carry no stance and count toward no figure. Anthropic's own blog and customer stories are left out, and an Anthropic engineer's workshop and two Anthropic partners' accounts are listed and not counted. Counts describe the sources reviewed and say nothing about SDK users at large. Product behavior is checked against Anthropic's documentation on the same dates.
Products in this piece
Sources
- Public source
- 27
- Official documentation
- 5
- Spotify Engineering: Coding Is No Longer the Constraint, on Honk and the Agent SDK (Jun 3, 2026)
- Sébastien Morel (Crystallize): How We Gave an AI Agent Our Codebase Without Giving It the Keys (Oct 13, 2026)
- Hacker News comment by Omnara's co-founder on replacing terminal parsing with the SDK (Feb 12, 2026)
- Hacker News comment by Terminal Use's co-founder on the SDK's hooks and permission protocol (Mar 9, 2026)
- Kyle Redelinghuys: The Claude Agent SDK, What It Is and Why It's Worth Understanding (Mar 2, 2026)
- Hacker News comment on a GitLab code reviewer built on the SDK (Feb 20, 2026)
- Hacker News comment on a user acceptance testing agent built on the SDK (Jul 2, 2026)
- Hacker News: Ask HN, Claude Agent SDK vs. Messages API for a production AI assistant (Apr 7, 2026)
- Ryan Daigle (Glide) on X: the SDK's subprocesses in a worker runtime (Apr 12, 2026)
- Claude Agent SDK TypeScript issue #97: session storage for Kubernetes deployments (Dec 11, 2025)
- Claude Agent SDK Python issue #555: resume silently starts a new session (Feb 7, 2026)
- Hacker News comment on rebuilding sessions, sandboxing, and approvals around the SDK for a multi-tenant web product (Apr 7, 2026)
- samael_voolt on X: a stable session ID restores prompt caching (Oct 8, 2026)
- Claude Agent SDK Python issue #974: bundled CLI never reads the prompt cache in headless runs (May 20, 2026)
- Claude Agent SDK TypeScript issue #93: cost tripled between 0.1.30 and 0.1.60 (Dec 6, 2025)
- Hacker News post on OpenHarness, written after finding the SDK too heavy to embed (Mar 31, 2026)
- Hacker News comment on a vendor-agnostic replacement for the SDK (May 1, 2026)
- validate.qa on X: dropping the SDK after spawn bugs in 0.2.139 (Jun 13, 2026)
- Claude Agent SDK TypeScript issue #296: musl binary chosen on glibc systems in 0.2.116 (Apr 21, 2026)
- Claude Agent SDK TypeScript issue #151: duplicate tool_use IDs in 0.2.19 (Jan 25, 2026)
- Dinmukhamed Mailibay (Laminar): How we instrumented Claude Agent SDK using a tiny Rust proxy (Dec 3, 2025)
- Elshad Karimov: Claude Agent SDK vs OpenAI Agents SDK, the same agent built in both (Jun 13, 2026)
- Thomson Reuters Innovation blog: CoCounsel Legal, Reimagined, rebuilt around the Claude Agent SDK (Mar 10, 2026)
- PostHog blog: What if your product built itself? (Jun 22, 2026)
- Claude Code issue #97074: headless SDK runs use about 1.8 times more of the 5-hour window (Sep 25, 2026)
- arthurain on X: a hit 5-hour quota returned as a successful result (Oct 1, 2026)
- Claude Agent SDK TypeScript issue #372: subprocess per session under 40 concurrent requests (Jul 9, 2026)
- Agent SDK docs: Overview (capabilities, login policy, license)
- Agent SDK docs: Hosting (subprocess model, session patterns, sizing)
- Agent SDK docs: Persist sessions to external storage
- Claude Help Center: Use the Claude Agent SDK with your Claude plan (updated Oct 7, 2026)
- Agent SDK docs: Track cost and usage