Claude Agent SDK Signals
What teams building AI features on Claude Agent SDK report, with every finding tied to its evidence.
Read the findingsWhat the evidence says
Practitioner sentiment, teams and leaders
+52 Strong positive from 22 sources that take a side
How this figure is madeStrong positive overall signal, net 52 points
From 21 authors in 22 discussions. Small sample. Build one pilot agent to measure the glue code and debugging time the framework adds.
- Analyzed
- 117 total sources
- Top praise
- Claude Code's Loop as a Library (11 of 11 sources) See praise evidence
- Top concern
- For teams and technical leaders, no concern theme reaches 5 sources from 3 authors.
Jun 2025 to Oct 2026.
98 sources from practitioners, by month and sentiment.
The 30 team and leadership accounts cover companies such as Spotify, PostHog and Crystallize and consultancies running client agents. Every account that discusses the harness praises it, and 6 report agents in daily use. Of the teams that describe a framework choice, nine kept the SDK, PostHog among them after two harnesses of its own. The CLI subprocess, Claude Code's defaults and plan billing are each raised by 3 sources, then token cost by 2.
Positive, mixed and negative sources
The score counts them as 12 positive, 8 mixed and 1 negative votes, one per author. See How the numbers work.
See the evidenceBelow: 2 praise themes.
Findings
Counted for Teams and technical leaders, from 30 sources. Change group
What practitioners praise
Every capability practitioners describe positively, ranked by how many sources say so. Open one to read who said it.
Claude Code's Loop as a LibraryAll 11 sources on it praise itBuilt-in tools, subagents, plan mode, compaction, hooks, and permissions come with the SDK, so builders skip writing the agent loop.
Counted for Teams and technical leaders.
Positive
Claire Vo, ChatPRD builder and How I AI host
What a harness is and how to build one with Claude Agent SDKPodcast episode, How I AI
Vo built a Sentry bug triage harness for ChatPRD with Claude Agent SDK and a terminal interface. It gathers evidence and creates follow-up artifacts so she can run a fixed workflow without rewriting a prompt.
Positive
vivekraja (Vivek), co-founder of Terminal Use (YC W26)
Comment on: Launch HN: Terminal Use (YC W26), Vercel for filesystem-based agentsHacker News comment
Hosting agents from several SDKs, finds the Claude Agent SDK's hooks and permission protocol the most granular, where other SDKs need a layer on top.
Positive
kmansm27 (Kartik), co-founder of Omnara (YC S25)
Comment on: Launch HN: Omnara (YC S25), Run Claude Code and Codex from anywhereHacker News comment
Omnara replaced terminal-output parsing with the SDK and its own UI, which the co-founder says is far easier to maintain.
Positive
Khaled Elfakharany, CTO and senior software engineer
I Found a Way to Use Claude Agent SDK Inside LangGraph NodesPersonal blog post
Runs each LangGraph node as an SDK call, so LangGraph controls the workflow while the SDK brings tools, context, and a model choice to each step.
Positive
vishnukool, founder of Agent37
Comment on: Show HN: Agent37, Monetize your Claude skills with shareable linksHacker News comment
Builds creators' agents on skills plus the SDK because, in the commenter's experience, it is the most capable agent available.
Positive
Shrivu Shankar, VP of AI at Abnormal AI
How I Use Every Claude Code FeaturePersonal newsletter (Substack)
His team is rolling out a planning tool built on the SDK in its work monorepo. He also uses the SDK to prototype agents, such as a threat investigation agent, before building full scaffolding.
Positive
Morgan Krey, Zed
Claude Code: Now in Beta in ZedCompany engineering blog
Zed wrapped Claude Code's SDK in an adapter that speaks the Agent Client Protocol, so Claude Code runs as its own process with Zed as the interface, and open-sourced the adapter.
Positive
Sébastien Morel, CTO of Crystallize
How We Gave an AI Agent Our Codebase Without Giving It the Keys?Company engineering blog
Crystallize's orchestrator starts one SDK session per Slack request, keeps credentials outside it, and routes heavy and light work to different Claude models.
"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."
Excerpt, Crystallize engineering Positive
Ian Vanagas, PostHog
What we wish we knew about building AI agentsCompany engineering blog
After two custom harnesses, PostHog rebuilt PostHog AI on the SDK with MCP tools and skills, which gave the agent a code sandbox for SQL and scripts. It advises teams to skip writing their own harness.
Positive
Balint Orosz, founder of Craft
Craft Agents - Our take on working effectively with Agents. (Available Open Source)Post on X
Craft built Craft Agents on the SDK after a CLI experiment; engineers use it as their main interface, and finance, marketing, support, and people teams use it too.
Positive
Shrivu Shankar, VP of AI at Abnormal AI
Building Multi-Agent Systems (Part 3)Personal newsletter (Substack)
For his recent agents he wraps Claude Code through the Agents SDK and keeps no custom loop. He finds the generic harness good enough for most use cases.
GILD's fact checks, which don't count as sources
Checked Oct 25, 2026.Anthropic documents built-in tools, hooks, subagents, MCP, permissions, sessions, skills, and plugins as Claude Code capabilities available in the SDK. Agent SDK overview
Open them in the recordAgents in Daily UseAll 6 sources on it praise itAgents built on the SDK that run in real work: code review, testing, migrations, Slack assistants, and grading.
Counted for Teams and technical leaders.
Positive
Sébastien Morel, CTO of Crystallize
How We Gave an AI Agent Our Codebase Without Giving It the Keys?Company engineering blog
Crystallize's orchestrator starts one SDK session per Slack request, keeps credentials outside it, and routes heavy and light work to different Claude models.
"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."
Excerpt, Crystallize engineering Positive
Josh Snyder, PostHog
What if your product built itself?Company engineering blog
PostHog runs the SDK in a sandbox for its research agent and for the agent that clones the repo, writes the fix, opens a PR, and keeps going until CI passes. It finds the agents do good work but cost a lot.
"Hand a generic report like "onboarding is broken" to the Agent SDK or Claude Code, and it'll go and change something regardless."
Excerpt, Josh Snyder and Cleo Lant Positive
Spotify Engineering, on a Code with Claude 2026 talk by Niklas Gustavsson, Chief Architect and VP of Engineering
Coding Is No Longer the Constraint: Scaling Developer Experience to Teams and Agents at SpotifyCompany engineering blog
Spotify runs its background coding agent Honk on the SDK inside its own harness, in Kubernetes pods that host many concurrent sessions.
"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."
Excerpt, Spotify Engineering Positive
Balint Orosz, founder of Craft
Craft Agents - Our take on working effectively with Agents. (Available Open Source)Post on X
Craft built Craft Agents on the SDK after a CLI experiment; engineers use it as their main interface, and finance, marketing, support, and people teams use it too.
Positive
Ian Vanagas, PostHog
What we wish we knew about building AI agentsCompany engineering blog
After two custom harnesses, PostHog rebuilt PostHog AI on the SDK with MCP tools and skills, which gave the agent a code sandbox for SQL and scripts. It advises teams to skip writing their own harness.
Positive
Shrivu Shankar, VP of AI at Abnormal AI
How I Use Every Claude Code FeaturePersonal newsletter (Substack)
His team is rolling out a planning tool built on the SDK in its work monorepo. He also uses the SDK to prototype agents, such as a threat investigation agent, before building full scaffolding.
What concerns them
Every concern practitioners raise, most raised first.
For teams and technical leaders, no concern reaches 5 sources from 3 authors.
Where the evidence agrees
Statements this group's sources agree on, then the ones close to agreement.
For teams and technical leaders, no statement reaches 5 sources from 3 authors.
Where it disagrees
Questions the sources split on, with the count on each side.
For teams and technical leaders, no question reaches 5 sources from 3 authors, with 2 on each side.
The evidence
Whose experience this is
The three groups, then every practitioner by the context they work in, and how emphasis changes with scale.
Three groups drawn from the same sources. Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. Signals open on Teams and technical leaders.
Teams and technical leaders
30 sources. Engineering leaders, founders, and company teams describe the agents they run on the SDK and the work they built around it.
Individual builders
68 sources. Individual builders describe agents they run, the sessions and caching they had to manage, and the billing terms they build under.
All practitioners
98 sources. These accounts combine company teams such as Spotify and PostHog with founders and solo developers. Their deployments range from one laptop to Kubernetes.
98 sources from practitioners, by the context they work in. Filled rows count for this group.
- Individual builderDevelopers building on the platform on their own68 sources
Emphasizes Sessions, caching, and the knobs the harness hides
- Small product team2 to 10 engineers, or an unnamed startup teamNo sources from this setting yet
- Product or platform team11 to 50 engineers, or a named company's team13 sources, counted here
Emphasizes Product agents in sandboxes and the hosting work around the subprocess
- Technical leadershipLeaders setting model and platform choices across teams15 sources, counted here
Emphasizes Internal and client agents built on the harness
- Organization-wideEnterprise or company-wide AI platform2 sources, counted here
Emphasizes A background coding agent on Kubernetes across the company, and a legal product rebuilt around the SDK
What changes with scale. Individual builders praise what the harness does and report the state, cost and plan-limit work it leaves them, including SDK runs using 1.7 to 1.8 times more of a plan's 5-hour window than Claude Code in October 2026. Named companies run it inside their own harness. Thomson Reuters rebuilt CoCounsel Legal around it, Spotify schedules many sessions in Kubernetes pods, Crystallize starts one session per Slack request, and PostHog runs it in sandboxes for its research and fix-writing agents. Teams building multi-tenant web products, 21st.dev among them, report rebuilding sessions, streaming, and approvals around it.
Where the sources come from
Where this group's sources were published.
Counted for Teams and technical leaders, from 30 sources. Change group
49 sources for this group, 19 of them documentation
- Product Documentation19, 39% of 49 sources
- Developer Communities15, 31% of 49 sources
- Engineering Blogs & Reviews13, 27% of 49 sources
- Podcasts, Talks & Videos2, 4% of 49 sources
Earlier evidence remains when it still applies. Records tied to closed issues, earlier billing terms, or the Claude Code SDK name carry Needs a recheck.
See sources by originThe record
Who each source is, how it knows, whether it still applies and what it counts toward.
Newest first, on its own page. Every count above opens its sources there.
Read all 117 sourcesAbout this read
What changed, and how we count
What changed
- Every record links to a public source, names its author and publication date, and stores the exact words it was coded from.
- A page shows a finding when at least 5 sources from 3 different authors discuss it. A disagreement also requires at least 2 sources on each side.
- Session, tracing, cost, and billing findings are checked against Anthropic's documentation and support pages as of October 25, 2026.
- Oct 26, 2026: adds company engineering posts and leader accounts from PostHog, Zed, Team 400, Autoolize, Craft, and Abnormal AI, plus LangChain's Applied Research lead.
- Oct 28, 2026: adds Operator Insights for Thomson Reuters' CoCounsel rebuild, Moon Banking's rollback, 21st.dev's switch from the Vercel AI SDK, and PostHog's pipeline, ledger quotes on 5 insights, and 33 records: 16 GitHub issues, 7 posts on X, 5 Reddit posts, 1 Hacker News comment, 2 podcast episodes, a conference workshop and a video. The 10 records from GILD's targeted search for problem reports, among them a hit 5-hour quota returned as a successful result, carry no stance and count toward no figure. Moves the How I AI episode to Podcasts, Talks & Videos and adds 14 measured results and the model releases records name.
- Oct 30, 2026: a second-model check removed the AI That Works workshop episode, which reports no deployment outcome, and Kowalah's video, an implementation consultancy's walkthrough of its own build. Moon Banking's tripled cost now reads as logged cost, which Anthropic traced to older versions undercounting while billing stayed correct, and its record no longer counts as a cost concern. Theo's X post is listed as a competitor's, Thariq Shihipar sits with individual builders, Thomson Reuters' podcast affiliation is unchecked, and the Claudius plan claim and the 18-agent fleet title say only what their pages say. Thariq Shihipar's Anthropic role reads as stated, since only the workshop page names it. 98 public sources remain.
Method
A running review of what teams building AI features on Claude Agent SDK report, from individual builders to company-wide platforms. Each finding links to its sources, and each conflict shows both sides.
Counts cover only the sources GILD reviewed.
Who counts
- Teams and technical leaders
- Small product team, Product or platform team, Technical leadership, Organization-wide
- Individual builders
- Individual builder
- All practitioners
- Everyone in the teams and leaders and individual builders groups
Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. GILD files each source in a group from how the account describes its author. A role or employer marked self-described is what the source states about themselves. Documentation checks facts for every group and never counts toward sentiment.
How the numbers work
- Overall signal. Positive votes minus negative ones, per 100 votes from the sources that take a side. Each author counts once, with their latest view, whoever they are and wherever they posted. The older weighted reading, where accounts GILD matched to a named person count three times when GILD lists the author among technical leaders and twice otherwise, is shown beside it as a second view. Where an author wrote more than one source, the line beside the split gives the votes, one per author, so the score follows from it. When one author's sources from the same day disagree, that author counts as mixed, and a half point rounds away from zero. Sources with no opinion are listed and not counted. Strong positive is +50 or higher, positive +20 to +49, mixed −19 to +19, negative −20 to −49, strong negative −50 or lower.
- Weighted reading, second view. Counted the older way, where an account GILD matched to a named person counts two or three times, teams and leaders read +67 strong positive (method sentiment/S1). GILD printed that number as "GILD score" until October 29, 2026; most of those matches rest on a name or a self-described role.
- Method change, Oct 30, 2026. GILD changed how practitioner sentiment counts, to one vote per author. Weighted, it read +71 Strong positive; with one vote per author it reads +58 Strong positive. No evidence changed.
- Unscored. A group gets a score only with 20 sources from 15 authors. Below that, one source moves the score about 5 points, so the read says "Unscored until 20 sources from 15 authors" and shows its counts.
- Independence. A source whose author works for the vendor, for one of its partners or programs, or for a competitor, and a page the vendor published or ran, is listed with that tie and counts toward nothing. One author counts on at most two records per product. The line under the score says how many authors and how many separate discussions the counted sources come from: comments in one thread are one discussion.
- Praise and concerns. How many of the sources that discuss a theme praise it or raise it.
- Consensus. A statement joins when 20% of the group's sources support it and fewer contradict it. Contradicting sources are counted separately. Below that line a statement shows as building. At 40% or more it reads Strong agreement; below, Recurring point.
- Disagreements. The count on each side. GILD names no winner and gives no share of all users.
- Minimum evidence. A finding needs 5 sources from 3 authors; a disagreement also needs 2 on each side. Smaller counts stay in the table below, marked under the minimum.
Labels on a source
A source's date and whether it still applies are separate: a recent post can describe an old version, and an old one can still hold.
- Documented
- Stated in official documentation.
- Reported
- A practitioner's account of their own use.
- Measured
- Backed by the source's own measurements, such as a comparative test or tracked team metrics.
- GILD analysis
- GILD's check of a claim against official documentation.
- Current
- Describes the product as it is now.
- Still applicable
- Older, and the behavior or lesson still holds.
- Needs a recheck
- May depend on a plan, version or setting that changed since it was written. It keeps counting until GILD rechecks it, since the rest of the account usually holds, and its card says so.
- Out of date
- No longer accurate. Kept in the record; counts toward nothing current.
Flagged for a recheck, of the sources that take a side: teams and leaders 2 of 22, individual builders 16 of 54, all practitioners 18 of 76.
Limits
- The counts cover 98 public sources: 27 Hacker News comments and posts, 33 GitHub issues, 15 posts on X, 5 Reddit posts, 15 company and personal blog posts, 2 podcast episodes and 1 conference workshop. They do not represent all SDK users. GitHub issues are bug reports and feature requests, which tilt the GitHub share toward problems. The 10 records from GILD's targeted search for problem reports carry no stance and count toward no figure. Two Reddit posts GILD found on October 28, 2026 had been deleted or removed and are left out. Anthropic's own blog and customer stories are left out, and an Anthropic engineer's workshop, Thomson Reuters' podcast episode, whose author's ties GILD has not checked, and an X post by the CEO of competing T3 Code and T3 Chat are listed and not counted.
- Sources run from the Claude Code SDK, before its September 2025 rename, to Python 0.2.x and TypeScript 0.3.x in October 2026. Each record keeps the version it states. Anthropic has since documented a SessionStore adapter for sessions and OpenTelemetry export. Closed issues on those topics carry Needs a recheck, as do billing accounts written before the October 7, 2026 plan update.
Every count, and the sources behind it
18 findings for 3 groups
Each row opens its sources for the group you are reading.
| Finding | Teams and technical leaders | Individual builders | All practitioners | Sources |
|---|---|---|---|---|
| Current readCurrent sentiment | 13 positive, 8 mixed, 1 negative | 15 positive, 34 mixed, 5 negative | 28 positive, 42 mixed, 6 negative | Sources |
| PraiseClaude Code's Loop as a Library | 11 of 11 positive | 10 of 10 positive | 21 of 21 positive | Sources |
| PraiseAgents in Daily Use | 6 of 6 positive | 7 of 7 positive | 13 of 13 positive | Sources |
| PraiseRun Cost & Plan Usage | 0 of 0 positive (under the minimum) | 5 of 6 positive | 5 of 6 positive | Sources |
| ConcernBundled CLI Subprocess | 3 of 3 raise it (under the minimum) | 9 of 9 raise it | 12 of 12 raise it | Sources |
| ConcernSessions on Local Disk | 1 of 1 raise it (under the minimum) | 9 of 9 raise it | 10 of 10 raise it | Sources |
| ConcernClaude Code's Defaults & Missing Knobs | 3 of 3 raise it (under the minimum) | 9 of 9 raise it | 12 of 12 raise it | Sources |
| ConcernToken Cost & Cache Misses | 2 of 2 raise it (under the minimum) | 6 of 6 raise it | 8 of 8 raise it | Sources |
| ConcernClaude Models Only | 0 of 0 raise it (under the minimum) | 5 of 5 raise it | 5 of 5 raise it | Sources |
| ConcernRegressions Between Releases | 1 of 1 raise it (under the minimum) | 7 of 7 raise it | 8 of 8 raise it | Sources |
| ConcernTracing Inside the Subprocess | 1 of 1 raise it (under the minimum) | 5 of 5 raise it | 6 of 6 raise it | Sources |
| ConcernPlan Billing & Terms Changes | 3 of 3 raise it (under the minimum) | 5 of 7 raise it | 8 of 10 raise it | Sources |
| ConsensusThe SDK saves writing the loop | 4 supporting (under the minimum) | 4 supporting (under the minimum) | 8 supporting (below threshold) | Sources |
| ConsensusShaped for one machine | 2 supporting, 1 contradicting (under the minimum) | 6 supporting (below threshold) | 8 supporting, 1 contradicting (below threshold) | Sources |
| ConsensusThe real work sits around it | 4 supporting (under the minimum) | 5 supporting (below threshold) | 9 supporting (below threshold) | Sources |
| DisagreementOnce a team has tried the SDK, does it keep it or write its own loop? | Kept the Agent SDK 9, Wrote or adopted another loop 1 (under the minimum) | Kept the Agent SDK 5, Wrote or adopted another loop 3 | Kept the Agent SDK 14, Wrote or adopted another loop 4 | Sources |
| DisagreementCan a builder rely on a Claude plan to run SDK agents? | The plan covers it 0, Terms keep shifting 2 (under the minimum) | The plan covers it 3, Terms keep shifting 5 | The plan covers it 3, Terms keep shifting 7 | Sources |
| Current readEvidence mix | 49 sources | 87 sources | 117 sources | Sources |
Fact checks
GILD's checks of claims against official documentation, kept apart from the source counts.
- Claude Code's Loop as a Library. Anthropic documents built-in tools, hooks, subagents, MCP, permissions, sessions, skills, and plugins as Claude Code capabilities available in the SDK. Agent SDK overview, checked Oct 25, 2026
- Run Cost & Plan Usage. Max 5x includes US$100 a month in API credits and Max 20x US$200, and the Agent SDK can spend them. Team seats add US$20 or US$100 each, pooled up to US$500, and unused credits expire each cycle. Monthly API credits for Max and Team plans, checked Oct 25, 2026
- Bundled CLI Subprocess. Anthropic documents one claude CLI subprocess per session, with a shell, working directory, and transcripts on local disk. It suggests 1 GiB RAM, 5 GiB disk, and 1 CPU per agent as a starting floor. Hosting the Agent SDK, checked Oct 25, 2026
- Sessions on Local Disk. A SessionStore adapter now mirrors session transcripts to an object store, key-value store, or database so another host can resume them. Memory files and working-directory artifacts still need their own storage. Persist sessions to external storage, checked Oct 25, 2026
- Claude Code's Defaults & Missing Knobs. Without a system prompt the SDK uses a minimal tool-calling prompt. The claude_code preset adds Claude Code's prompt, and a fully custom prompt replaces both and takes over tool guidance and safety rules. Modifying system prompts, checked Oct 25, 2026
- Token Cost & Cache Misses. The SDK's total_cost_usd and costUSD fields are client-side estimates from a bundled price table. For billing figures, Anthropic points to the Usage and Cost API or the Console. Track cost and usage, checked Oct 25, 2026
- Claude Models Only. The SDK authenticates with an Anthropic API key or with Claude on Amazon Bedrock, Claude Platform on AWS, Google Cloud's Agent Platform, or Microsoft Foundry. Agent SDK quickstart, checked Oct 25, 2026
- Claude Models Only. The model option takes a Claude alias or full model name. Left unset, the SDK runs Claude Code's default model, and a fallbackModel chain takes over when the primary model fails. Agent SDK TypeScript reference, checked Oct 28, 2026
- Regressions Between Releases. PyPI lists 139 releases of the Python package in the twelve months to October 25, 2026, the latest 0.2.174 on October 24. claude-agent-sdk on PyPI, checked Oct 25, 2026
- Tracing Inside the Subprocess. Anthropic documents OpenTelemetry export of metrics, log events, and traces from the CLI subprocess. The SDK emits no telemetry itself, and traces require a beta flag. Observability with OpenTelemetry, checked Oct 25, 2026
- Plan Billing & Terms Changes. Since October 7, 2026, Max and Team plans include monthly API credits that cover the Agent SDK. SDK use can still draw on subscription limits, and the June 15, 2026 change was paused. Use the Claude Agent SDK with your Claude plan, checked Oct 25, 2026