| name | claude-api |
| description | Build and debug Anthropic Managed Agents in TypeScript. TRIGGER when: code imports `@anthropic-ai/sdk`; user asks about Managed Agents, agent sessions, agent environments, agent tools, agent events, agent memory, or the Anthropic SDK; user creates/modifies agents, sessions, environments, vaults, memory stores, or skills via the Anthropic API. SKIP: file imports `openai`/other-provider SDK, provider-neutral code, general programming/ML. |
| license | Complete terms in LICENSE.txt |
Managed Agents — TypeScript
This skill helps you build Managed Agents with the Anthropic TypeScript SDK (@anthropic-ai/sdk). Managed Agents are server-managed stateful agents where Anthropic runs the agent loop and hosts a per-session container for tool execution (bash, file ops, code execution).
Before You Start
Scan the target file for non-Anthropic provider markers — import openai, OpenAI(, gpt-4, gpt-5. If found, stop and ask the user whether they want to switch to Claude.
Defaults
Unless the user requests otherwise, use:
- Model:
claude-opus-4-7 (exact string — do not append date suffixes)
- Thinking:
thinking: {type: "adaptive"} for anything remotely complicated
- Effort:
output_config: {effort: "xhigh"} for agentic workloads on Opus 4.7
CRITICAL: Use only exact model ID strings. claude-opus-4-7, claude-opus-4-6, claude-sonnet-4-6, claude-haiku-4-5. Do not append date suffixes. If the user requests an older model, read shared/models.md.
SDK Usage
Never guess SDK usage. Function names, class names, method signatures, and import paths must come from typescript/managed-agents/README.md or the official SDK repo (URL in shared/live-sources.md). If a binding isn't documented, WebFetch it — don't infer.
Don't reimplement SDK functionality. Use typed exception classes (Anthropic.RateLimitError, etc.) and SDK types (Anthropic.MessageParam, Anthropic.Tool, Anthropic.Message, etc.) instead of redefining equivalent interfaces.
Managed Agents Overview
Mandatory flow: Agent (once) → Session (every run).
model/system/tools live on the Agent, never the session.- Agents are persistent — create once, reference by ID. Store the agent ID returned by
agents.create and pass it to every subsequent sessions.create; do not call agents.create in the request path.
Beta headers: managed-agents-2026-04-01 — the SDK sets this automatically for all client.beta.{agents,environments,sessions,vaults,memory_stores}.* calls.
First-party only. Not available on Amazon Bedrock, Google Vertex AI, or Microsoft Foundry.
Subcommands
| Subcommand | Action |
|---|
managed-agents-onboard | Walk the user through setting up a Managed Agent from scratch. Read shared/managed-agents-onboarding.md immediately and follow its interview script. |
Current Models (cached: 2026-04-15)
| Model | Model ID | Context | Input $/1M | Output $/1M |
|---|
| Claude Opus 4.7 | claude-opus-4-7 | 1M | $5.00 | $25.00 |
| Claude Opus 4.6 | claude-opus-4-6 | 1M | $5.00 | $25.00 |
| Claude Sonnet 4.6 | claude-sonnet-4-6 | 1M | $3.00 | $15.00 |
| Claude Haiku 4.5 | claude-haiku-4-5 | 200K | $1.00 | $5.00 |
Thinking & Effort (Quick Reference)
Opus 4.7 — Adaptive only. thinking: {type: "adaptive"}. budget_tokens returns 400. Sampling params (temperature, top_p, top_k) also removed. See shared/model-migration.md.
Effort parameter (GA): output_config: {effort: "low"|"medium"|"high"|"xhigh"|"max"} (inside output_config). xhigh is the sweet spot for agentic workloads on Opus 4.7. max is Opus-only.
Thinking content omitted by default on 4.7: Set thinking: {type: "adaptive", display: "summarized"} to restore visible progress.
Task Budgets (beta, Opus 4.7): output_config: {task_budget: {type: "tokens", total: N}} — minimum 20,000; beta header task-budgets-2026-03-13. Distinct from max_tokens. See shared/model-migration.md.
Reading Guide
Getting started / setting up a new agent:
→ Read typescript/managed-agents/README.md + shared/managed-agents-overview.md + shared/managed-agents-onboarding.md
Core concepts (agent lifecycle, sessions, containers):
→ Read shared/managed-agents-core.md
Environments (sandboxes, file mounts, secrets):
→ Read shared/managed-agents-environments.md
Tools (built-in tools, custom tools, skills, MCP):
→ Read shared/managed-agents-tools.md
Events & streaming (SSE event types, stream handling):
→ Read shared/managed-agents-events.md
Session outcomes (success, failure, idle, terminated):
→ Read shared/managed-agents-outcomes.md
Client patterns (reconnect, interrupt, queued/processed gate, file mounts):
→ Read shared/managed-agents-client-patterns.md
Multi-agent orchestration:
→ Read shared/managed-agents-multiagent.md
Webhooks:
→ Read shared/managed-agents-webhooks.md
Memory stores:
→ Read shared/managed-agents-memory.md
Full API reference:
→ Read shared/managed-agents-api-reference.md
Model migration / upgrading:
→ Read shared/model-migration.md
Error codes:
→ Read shared/error-codes.md
Latest docs URLs (for WebFetch):
→ Read shared/live-sources.md
Model capabilities / context windows:
→ Read shared/models.md
Common Pitfalls
- Agents are persistent. Don't call
agents.create in the request path — create once, store the ID, reference it in sessions.create.
model/system/tools go on the Agent, not the Session.- Opus 4.7 thinking: Adaptive only.
budget_tokens returns 400.
- 4.6/4.7 prefill removed: Assistant message prefills return 400 on Opus 4.6/4.7/Sonnet 4.6.
- Confirm migration scope before editing: When asked to migrate code without naming specific files, ask first.
max_tokens defaults: Non-streaming: ~16000. Streaming: ~64000.- Don't truncate inputs. If content is too long, notify the user.
