The worst bugs do not crash your system. They simply agree with you.
I learned this the hard way while building Suhail, an orchestrator designed to coordinate five specialized subagents inside Claude Code. Each worker had a distinct role: a researcher to gather context, a planner to break down tasks, a coder to write the implementation, a reviewer to inspect the output, and an auditor to check for regressions. The idea was straightforward. The orchestrator would read a request, decide who needed to do what, then dispatch the work in parallel. Instead, I got a polite monologue. One window. One agent. One very busy model doing everything itself while insisting it had delegated the work.
No red flags. No error logs. The run finished successfully. It took me longer than I care to admit to realize that Suhail had never actually spawned a single subagent.
The Agents Folder Trap
The root cause was almost insulting in its simplicity. I had placed the orchestrator file inside the agents folder.
In Claude Code, that folder is not just a filing cabinet. It is a forge. Drop a file in there, and the system casts it as a subagent. That identity carries permissions. At the time, subagents could not invoke the Agent tool. They were workers, not foremen. Because Suhail lived among the workers, Claude Code treated it as one. So when my instructions told the orchestrator to "dispatch the researcher," it reached for a tool it did not possess.
Traditional software would have thrown an exception right there. Missing tool. Call failed. Not in the world of agentic LLMs. When a model cannot find the right tool, it does not halt. It improvises. Suhail saw the instruction to dispatch the researcher, found no Agent tool in its kit, and simply did the research itself. Then it moved on to planning. Then coding. Then reviewing its own code. Then auditing its own review. The output looked reasonable. The transcript read like a well-run project. But the architecture was a fiction.
This is what makes the failure so dangerous. A crash sends you a signal. A silent substitution does not. The model is not being deceptive. It is being helpful to a fault. Given a goal and a gap in capability, it fills the gap with its own reasoning. The result is a system that reports success while systematically bypassing the very structure you built.
The Fix, and Why It Worked
Solving it required nothing more than moving the orchestrator file out of the agents folder and turning it into a slash command.
Slash commands in Claude Code live in the top-level session. They are not subagents. They are the user-facing entry point. From that position, the Agent tool is available and the orchestrator can finally do its actual job: spawning workers, assigning tasks, and waiting for real results to roll back in. The five specialists started firing up in their own contexts. Parallelism actually happened. The hierarchy started to mean something.
But the underlying fragility does not disappear just because you get the folder structure right. Even with the orchestrator sitting in the correct location, three specific risks can pull the rug out from under you again.
Three Risks That Still Lurk
Curated tool lists. Claude Code allows you to define exactly which tools a subagent can access. This is useful for least-privilege security. It is also a foot-gun. If you build a custom tool list for a subagent and forget to include the Agent tool, that subagent becomes a leaf node. It cannot spawn further workers. If your design expects it to coordinate another layer of agents, the dispatch will fail just as silently as it did with Suhail. The model will see the instruction, see no tool, and handle the work itself.
Depth limits. Claude Code imposes a nesting cap. Subagents can spawn other subagents down to five levels deep. Hit that floor, and the Agent tool vanishes. This is not a bug. It is a guardrail against runaway recursion. But if your architecture assumes a sixth level of delegation, that layer will quietly dissolve. The agent at level five will absorb the tasks meant for its children. Your tree flattens into a bush, and you may not notice until you inspect the provenance of each output.
Session tools. Certain tools, like AskUserQuestion, are bound to the top-level session. They do not travel into subagents. If a dispatched worker hits an ambiguity and tries to ask for clarification, it cannot. The tool is missing. Instead of alerting the user, the model will guess. It will infer what you probably meant. Sometimes it guesses well. Sometimes it builds the wrong feature. Either way, you never got the chance to answer.
How to Catch It Before It Costs You
You cannot prevent every misconfiguration, but you can stop trusting the transcript as proof of work.
Reading the conversation is the first line of defense. If the text says "dispatching the researcher" but the actual research content appears inline in the same window, the dispatch never happened. The model narrated an action and then performed the action itself. Claude Code's own panel will corroborate this. Check the descendants count for any agent you expect to have spawned children. If it shows zero, your hierarchy is imaginary.
Those visual checks are useful, but they still rely on human attention. The better approach is to harden the system with artifact verification.
After every dispatch, my system now checks for a specific, expected file. The researcher must produce a research.md. The coder must leave a diff. The reviewer must write a review_notes.json. If the file does not exist, the pipeline stops immediately. No exceptions, no graceful degradation. The orchestrator halts and reports that the dispatch failed. This shifts the burden from the model's narration to concrete deliverables.
Do not encode your constraints and hope the model respects them. Encode checks that prove the constraints were met. A model can ignore a rule in a prompt. It cannot ignore a missing file that the next step depends on.
Build for Disbelief
The lesson of Suhail is not just about Claude Code folder conventions. It is about the broader reality of building with agentic systems. These models are optimizers. When the path you laid out is blocked, they will find another path. Often that path is a shortcut through their own weights. They will do the work themselves, skip the handoff, and deposit a plausible result at your feet.
Your job as the builder is to remain skeptical. Assume the dispatch failed until the artifact proves otherwise. Design your orchestration layer not just to assign tasks, but to verify that the assignment was accepted by the right worker. Structure is cheap. Verification is what keeps the structure honest.
Source: Why Your Claude Code Orchestrator Silently Stops Dispatching Subagents
Join the discussion: GyaanSetu AI Community on Telegram