25 guides, updated Oct 15, 2026 · status re-checked
Your agent broke. Find the fix that worked.
A searchable board of real AI agent failures and the fixes that worked. Each guide gives the cause, numbered steps, and a status backed by a linked upstream source.
Pick a framework
Search guides
- AutoGenFixed (verified)
'str' object has no attribute 'get' when the chat finishes
The task output prints, then the run crashes with AttributeError: 'str' object has no attribute 'get'.
Last checked#schema#error-messages3 steps → - LangGraphUnverified
"tool_use ids were found without tool_result blocks" on a saved thread
An agent with a database checkpointer fails on every new message in one thread with a tool_use/tool_result mismatch error.
Last checked#state#tool-calls3 steps → - AutoGenUnverified
Adding a second agent to the quick-start makes it loop forever
After registering one more agent, the program never ends.
Last checked#messaging#loop3 steps → - Keystone slot · sponsorYour tool fits hereOne tool, every page, one week.$99/week · 0 sold
- LangGraphWorkaround
Agent repeats the same tool call until it hits the recursion limit
A prebuilt agent keeps calling tools step after step and only stops when GraphRecursionError is raised.
Last checked#loop#tool-calls5 steps → - n8nUnverified
AI Agent tools fail with 'schema must be type object, got None' after upgrading
After upgrading, the agent fails before calling any tool.
Last checked#regression#tool-calls3 steps → - browser-useFixed (verified)
Browser agent can't pick an option from a dropdown
The agent clicks the dropdown, then fails or loops trying to choose an option.
Last checked#browser#regression4 steps → - browser-useUnverified
Browser agent keeps emitting empty actions until max steps
Sometimes the agent outputs an empty action every step until the step limit, even though it read the right data.
Last checked#schema#tool-calls3 steps → - browser-useUnverified
Browser agent keeps going after it already called done
In headless mode with a custom structured-output controller, the run never ends after the task is complete.
Last checked#loop#schema3 steps → - LangGraphOpen
Cancelled async tool leaves history invalid even with handle_tool_errors=True
After a tool task is cancelled, the next model call fails with INVALID_CHAT_HISTORY.
Last checked#state#tool-calls3 steps → - browser-useOpen
Click on a missing element is reported as success
Actions continue as if a click or dropdown pick worked, but nothing happened on the page.
Last checked#error-messages3 steps → - AiderUnverified
Coding agent stuck on 'did not conform to the edit format'
Every edit fails with 'SEARCH/REPLACE block failed to match' and the session can't continue.
Last checked#parsing#loop3 steps → - Hermes AgentWorkaround
Error messages name the symptom, not the cause
You fix what the error says, and it still fails. For example, a 'payment/credit error' when there is simply no API key.
Last checked#error-messages3 steps → - CrewAIWorkaround
Every tool call arrives with empty arguments and fails validation
Each tool call fails with 'Field required ... input_value={}' and the agent repeats until max iterations.
Last checked#tool-calls#schema3 steps → - MCPFixed (verified)
Fetch tool can be steered to localhost or cloud metadata
An agent with the fetch server can be told, or redirected, to read internal addresses.
Last checked#security3 steps → - MCPFixed (verified)
Filesystem server says 'path outside allowed directories' on the first call
The first tool call right after connecting is denied. A moment later the same call works.
Last checked#race#setup2 steps → - AutoGenUnverified
Graph flow errors with 'unable to find next speaker' when a loop has tools
A graph with a reviewer→writer revise loop fails at runtime even though the edge exists.
Last checked#loop3 steps → - LangChainWorkaround
JSON output parser returns None instead of raising on bad JSON
A chain returns None for some malformed outputs, and later code fails far from the cause.
Last checked#parsing#error-messages3 steps → - MCPUnverified
npx-based server shows zero tools when Node comes from nvm
The server is configured but never connects. Servers built as .exe or binaries work.
Last checked#setup3 steps → - n8nOpen
One bad tool call ends the whole agent run
The run ends with 'Received tool input did not match expected schema' and the user gets nothing.
Last checked#tool-calls#schema3 steps → - LangChainWorkaround
Parser fails when the model wraps JSON in ```JSON instead of ```json
parse_json_markdown raises on output that is valid JSON inside a code fence.
Last checked#parsing2 steps → - LangGraphWorkaround
Setting recursion_limit to 10000 silently does nothing
with_config({"recursion_limit": 10000}) has no effect on a subgraph or agent.
Last checked#config3 steps → - MCP Python SDKWorkaround
Stateless MCP server logs ERROR every time a client disconnects
Two ERROR records appear whenever a client cancels a request, and alerts fire.
Last checked#error-messages#transport3 steps → - MCP Python SDKOpen
Streamable HTTP client keeps reconnecting past its retry limit
A long-running request never resolves. The client keeps reconnecting.
Last checked#transport2 steps → - LangGraphUnverified
Token limit exceeded because retriever output piles up in chat history
A ReAct agent with a persistent checkpointer starts failing with token-limit errors after a few turns.
Last checked#state4 steps → - ContinueWorkaround
Tool calls fail because arguments come back as XML-like text
In agent mode, tool calls break with empty arguments or list/index errors when using a self-hosted compatible endpoint.
Last checked#tool-calls#serving3 steps →
That's all 25. Missing one? Ask on the board.
What each status means
- Open
The upstream issue is still open. The steps are the best known path for now.
- Workaround
Upstream closed it as not planned, stale or duplicate. The steps get you around it.
- Fixed
A merged PR or commit fixed it upstream. The guide links that evidence.
Featured fixes
How featuring works →FAQ
Who writes the fixes?
A bot gathers failures from public reports and issue trackers, then drafts the fix in its own words, with a link to every source.
What does Unverified mean?
No proof yet that a fix shipped upstream. We never mark a guide Fixed without a linked change.
Is it free?
Searching and reading every fix is free.
Does featured mean paid?
Yes, and it's labeled. Featured fixes still have to work.
What do you do with my email?
We use it only to send the Weekly Fix, weekly on Mondays, starting soon. Nothing else. Never sold.