AgentGuides Update Report — April 18, 2026

AgentGuides Update Report — April 18, 2026

Prepared by: Automated review process
Review date: April 18, 2026
Previous review: April 17, 2026
Scope: All 18+ framework guides across Python, TypeScript, .NET, and Go

---

Executive Summary

This report documents the changes made during the April 18, 2026 review of the AgentGuides repository. The primary driver was a same-day release of new versions of the Anthropic Claude Agent SDK (Python v0.1.63, TypeScript v0.2.113) with developer-quality-of-life improvements. All other framework guides were audited and confirmed current; the PydanticAI comprehensive guide received a documentation enhancement for the v1.84.0 features released the previous day.

---

Changes Made

1. Anthropic Claude Agent SDK (Python) — Updated to v0.1.63

File: Anthropic_Claude_Agent_SDK_Guide/README.md
File: Anthropic_Claude_Agent_SDK_Guide/anthropic_claude_agent_sdk_comprehensive_guide.md

New features documented:

get_context_usage() on ClaudeSDKClient

The new get_context_usage() method returns a per-category breakdown of context window consumption:

• system_tokens — tokens consumed by the system prompt
• conversation_tokens — tokens from conversation history
• tool_tokens — tokens attributed to tool results
• total_tokens / context_window / utilisation_pct — aggregate stats

This enables proactive context management — for example, triggering compaction or summary when utilisation exceeds a threshold.

typing.Annotated support in @tool and create_sdk_mcp_server

Tool function parameters can now carry per-parameter descriptions using Annotated[type, "description"]. The SDK automatically generates accurate JSON Schema from these annotations, improving LLM tool-call accuracy without requiring a separate schema dictionary.

from typing import Annotated
from claude_agent_sdk import tool

@tool("search_codebase", "Search the project codebase for patterns")
async def search_codebase(
    pattern:   Annotated[str,  "The regex or literal string to search for"],
    directory: Annotated[str,  "Relative path to the directory to search in"] = ".",
    max_results: Annotated[int, "Maximum number of results to return (1–100)"] = 20
) -> dict:
    ...

tool_use_id and agent_id in ToolPermissionContext

Permission handler callbacks now receive two additional context fields:

• tool_use_id — the unique ID of the specific tool call (allows distinguishing parallel invocations)
• agent_id — the ID of the subagent making the permission request (useful in multi-agent setups)

This makes it possible to build fine-grained permission policies that behave differently based on which agent is requesting permission and which specific parallel invocation is being evaluated.

---

2. Anthropic Claude Agent SDK (TypeScript) — Updated to v0.2.113

File: Anthropic_Claude_Agent_SDK_TypeScript_Guide/README.md

Same features as the Python SDK, with TypeScript-idiomatic naming:

• getContextUsage() (camelCase) returning ContextUsage with systemTokens, conversationTokens, toolTokens, totalTokens, contextWindow, utilisationPct
• JSDoc @param annotations on @tool parameters for automatic JSON Schema description generation
• toolUseId and agentId in ToolPermissionContext

---

3. PydanticAI Comprehensive Guide — v1.84.0 Features Completed

File: PydanticAI_Guide/pydantic_ai_comprehensive_guide.md

The revision history entry for v1.84.0 was incomplete (only mentioned the Google FileSearchTool regex fix). The following new sections and code examples were added:

Feature	Description
OllamaModel subclass	Dedicated first-class class replacing OpenAIModel workaround; fixes structured output on Ollama Cloud
XSearchTool / FileSearchTool for xAI	Built-in search and file retrieval tools for Grok provider
FastMCPToolset metadata injection	inject_metadata=True attaches call_id, timestamp, and agent_id per tool call
Bedrock prompt cache TTL	cache_ttl parameter on BedrockModel for configurable response caching
Stateful OpenAICompaction	mode='stateful' compacts history while preserving state references
Claude Opus 4.7 support	anthropic:claude-opus-4-7 recognised as a valid model string

---

4. README.md — Stale Date Corrected

File: README.md

The “Last Updated” section incorrectly showed “November 21, 2025” despite the guides having been updated through April 2026. This has been corrected to “April 18, 2026” with per-framework version details. Revision history updated.

---

5. versions.json — Updated

File: versions.json

Field	Previous	Updated
last_updated	April 17, 2026	April 18, 2026
anthropic-claude-agent-sdk-python	0.1.60 (April 16, 2026)	0.1.63 (April 18, 2026)
anthropic-claude-agent-sdk-typescript	0.2.110 (April 16, 2026)	0.2.113 (April 18, 2026)
anthropic-claude-sdk notes	v0.1.60 features	Extended to include v0.1.63 features

---

Framework Audit Summary

All 18+ framework guides were audited against official release notes, changelogs, and PyPI/npm registries. The table below summarises the findings:

Framework	Guide Version	Latest Release	Status
OpenAI Agents SDK (Python)	0.14.1	0.14.1 (Apr 15)	✅ Current
OpenAI Agents SDK (TypeScript)	0.8.3	0.8.3 (Apr 9)	✅ Current
PydanticAI	1.84.0	1.84.0 (Apr 17)	✅ Current (enhanced)
CrewAI	1.14.0	1.14.0 (Apr 7)	✅ Current
AG2 (AutoGen)	0.11.5	0.11.5 (Apr 5)	✅ Current
LangGraph (Python)	1.1.6	1.1.6 (Apr 10)	✅ Current
LangGraph (TypeScript)	1.2.8	1.2.8 (Apr 11)	✅ Current
LlamaIndex (Python)	0.14.20	0.14.20 (Apr 3)	✅ Current
LlamaIndex (TypeScript)	@llamaindex/workflow 1.1.4	1.1.4 (Apr 15)	✅ Current
Haystack	2.27.0	2.27.0 (Apr 1)	✅ Current
Google ADK (Python)	1.30.0	1.30.0 (Apr 13)	✅ Current
Google ADK (Go)	1.0.0 GA	1.0.0 GA (Apr 8)	✅ Current
Google ADK (TypeScript)	0.1.x	0.1.x (Dec 2025)	✅ Current
Amazon Bedrock Strands	1.35.0	1.35.0	✅ Current
Microsoft Agent Framework	1.0 GA	1.0 GA (Apr 3)	✅ Current
Semantic Kernel (Python)	1.41.2	1.41.2 (Apr 8)	✅ Current
Semantic Kernel (.NET)	1.74.0	1.74.0	✅ Current
Mistral Agents API	2.0.1	2.0.1 (Mar 12)	✅ Current
Anthropic Claude SDK (Python)	0.1.60 → 0.1.63	0.1.63 (Apr 18)	✅ Updated
Anthropic Claude SDK (TypeScript)	0.2.110 → 0.2.113	0.2.113 (Apr 18)	✅ Updated
SmolAgents	1.24.0	1.24.0 (Jan 16)	✅ Current
AutoGen (Legacy)	—	Deprecated	✅ No change needed

---

No Action Required

The following items were reviewed and found to require no changes:

• Deprecated methods: No deprecated patterns found in active guides. The AutoGen Legacy guide correctly directs users to AG2.
• Breaking changes: All breaking changes from recent releases (OpenAI SDK v2 requirement, LlamaIndex hard-removed legacy agents, Haystack Python 3.9 drop, PydanticAI result_* removal) are documented in their respective guides.
• Code examples: All guides have comprehensive code examples covering basic use through advanced patterns including multi-agent systems, tool use, memory, context management, checkpoints, observability, and streaming.

---

Cross-Cutting Standards Verification

Standard	Status
British English spelling	✅ Consistent throughout
Revision history in all guides	✅ Present and updated
Version numbers in guide headers	✅ All current
Breaking changes documented	✅ All documented
Code examples for new features	✅ Added for v0.1.63 and v1.84.0
Beginner-to-advanced learning paths	✅ Present in all major guides

---

Recommended Follow-Up (Next Review)

1. Google ADK Python — The bi-weekly release cadence means a new version (approximately 1.32.x) is likely within 2 weeks. Monitor for new session management features and additional language model support.
2. LangGraph — Watch for v1.2.x with potential additional v2 API enhancements.
3. Haystack — v2.28.0 expected within weeks; monitor for additional component integrations.
4. PydanticAI — Rapid release cadence (2–3 releases per week); monitor for v1.85+.
5. CrewAI — Monitor for v1.15.0 and any changes to the Flows system.

---

Report generated: April 18, 2026