Chapter 11 Troubleshooting

Common problems and how to fix them

This chapter covers the most common issues you’ll encounter when building and running Agentish workflows, along with their solutions.

Export & Compilation Errors

These errors appear when you click “Download Bundle” or when the sandbox compiles your ASL. They block execution until fixed.

Error	Cause	Fix
“Graph requires an Entry Point node.”	No Entry Point on the canvas.	Right-click → add an Entry Point node.
“Entry node has no outgoing edges.”	Entry Point is disconnected.	Connect the Entry Point’s output to your first LLM or Router node.
“No terminal LLM node found.”	Every LLM Node has an outgoing edge — execution can never reach END.	Ensure at least one LLM Node has its output slot unconnected.
“Node ‘X’ has no incoming flow edge and is unreachable.”	An LLM or Router node is floating on the canvas with no incoming connections.	Connect it to the graph or delete it if unused.
“Worker node ‘X’ is not connected to any LLM node.”	A Worker exists but isn’t connected as a tool to any LLM Node.	Draw an edge from an LLM Node to the Worker, or delete the Worker.
“Router node ‘X’ has N outgoing flow edge(s). Routers must have at least 2.”	A Router has fewer than 2 outputs.	Add more Router Values and connect additional output edges.
“LLM node ‘X’ has multiple incoming edges but ‘loop_mode’ is not configured.”	A loop target LLM Node doesn’t have loop_mode set.	Click the node, set Loop mode to “Fresh” or “Continue” in the inspector.
“Infinite loop detected: cycle [...] contains no Router node.”	A cycle exists with no Router to control iteration.	Add a Router Node in the cycle to decide when to exit.
“Infinite loop detected: cycle [...] contains a Router but all its outgoing edges stay within the cycle.”	The Router in a cycle has no exit path.	Add an output edge from the Router to a node outside the cycle.

Infinite Loops at Runtime

Even with valid topology, your agent can loop forever at runtime if the Router never chooses the exit path. Signs of a runtime infinite loop:

The sandbox execution runs for a very long time (approaching the 10-minute timeout).
The count variable in the final state is unusually high.
Langfuse shows the same nodes executing repeatedly with similar outputs.

How to Fix

Improve the Router prompt. Add specific, measurable exit criteria. Example: “After 3 iterations, always route to Finalizer.”
Add a feedback state variable. Give the loop target LLM new context on each iteration so it doesn’t produce the same output.
Lower ambiguity. If the Router prompt says “route to Finalizer when done,” the LLM may never think it’s “done.” Be concrete: “route to Finalizer when analysis_result is over 200 words and addresses all three areas.”

Recursion Limits

LangGraph has a built-in recursion limit that caps the total number of node executions in a single run. If your workflow exceeds this limit, execution stops with a recursion error.

This typically happens with loops that iterate too many times or deeply nested worker calls. The fix is the same as for infinite loops: improve exit conditions and Router prompts.

Tool Iteration Limit Reached

If the LLM hits max_tool_iterations before completing its task, it’s forced to produce a final response. The output may be incomplete.

How to Fix

Increase the limit if the task genuinely requires many tool calls (e.g., investigating 20+ functions in a binary).
Improve the system prompt to make the LLM more efficient. Tell it to prioritize the most promising leads instead of exhaustively checking everything.
Customize the warning message to guide the LLM on how to wrap up: “Focus on the functions most likely to contain the flag. Produce your best answer now.”
Split the work across multiple nodes or Worker nodes, each with its own tool budget.

Improving Prompt Quality

Most workflow failures trace back to prompt quality. The LLM does what you tell it — if the prompt is vague, the output will be vague.

Common Prompt Problems

Problem	Symptom	Fix
Too vague	LLM produces generic, unhelpful output.	Add specific instructions: what to analyze, what format to use, what to avoid.
Too long	LLM ignores parts of the prompt or gets confused.	Trim to essentials. Use bullet points instead of paragraphs.
Conflicting instructions	LLM oscillates or produces inconsistent output.	Review for contradictions. Prioritize instructions clearly.
Missing output format	LLM returns prose when you need structured data.	Add an output format section. Use structured output schema.
No constraints	LLM hallucinates, guesses, or makes assumptions.	Add explicit constraints: “Never guess. Use tools to verify.”

Tip: Write and iterate on prompts in an external text editor with more screen space, then paste into the inspector. The small textarea in the Agentish inspector is not ideal for long prompts. Try notepad.link or any plain text editor.

LLM Picks the Wrong Tool

Too many tools assigned. Reduce to 2–5 per node.
Tool descriptions are ambiguous. The LLM reads tool descriptions to decide which to call. If two tools sound similar, the LLM may pick the wrong one. (Note: MCP tool descriptions come from the challenge and can’t be edited, but you can guide the LLM via system prompt.)
System prompt doesn’t mention tools. Tell the LLM which tools to use and in what order. Example: “First use list_functions to get an overview, then use decompile_function on the most promising targets.”

State Variable Issues

Problem	Cause	Fix
Downstream node sees empty variable	The upstream node didn’t select the variable in “Global state to update.”	Check that the upstream node’s output_state_keys includes the variable.
Variable has wrong value	Two nodes write to the same variable (last-write-wins).	Use separate variables for separate nodes.
Template variable shows raw `{name}`	Variable name in the prompt doesn’t match a state variable.	Check spelling. Variable names are case-sensitive in state schema (lowercased).

Execution Timeout

The sandbox has a 10-minute timeout per execution. If your agent exceeds this:

Check for infinite loops (see above).
Reduce max_tool_iterations on nodes that call many tools.
Simplify the workflow — fewer nodes, fewer loop iterations.
Use a faster model if available.

Chapter Summary

Key Takeaways:

Export errors are caught by the topology validator — read the error message carefully, it tells you exactly what to fix.
Runtime infinite loops come from vague Router exit criteria — make them specific and measurable.
Most failures trace back to prompt quality. Write detailed, structured prompts in an external editor.
Keep tools to 2–5 per node and guide the LLM on which tools to use in the system prompt.
Read the final state and Langfuse traces to debug — trace problems backwards from the wrong output.

← Chapter 10: Logging