I’m working on building a multi-agent system using a supervisor pattern. Each individual agent works perfectly when I test them separately. They call the right functions and give me exactly what I expect. But when I put them under the supervisor control, weird things start happening.
First problem is with tool execution visibility. When I run agents by themselves, I can see all the function calls happening in my notebook (usually 3-4 calls per agent). But with the supervisor running things, I only see one tool execution in the interface. The strange part is that when I check the LangSmith traces, all the tools are actually being called and working correctly.
Second issue is about the final results. Individual agents return their complete detailed output when run alone. However, under supervisor management, I’m getting shortened summaries instead of the full results from the last agent. Again, LangSmith shows the complete outputs are being generated internally.
Anyone else run into similar behavior with supervisor setups? Would really appreciate any tips on how to debug this or if there are known workarounds.
Had this exact problem last month when I built my own supervisor setup. The supervisor basically sits between your agents and the output, processing and sometimes filtering what gets through. When agents run solo, they output directly - but with supervision, everything goes through that intermediary layer. For the tool visibility issue, the supervisor usually bundles all the tool calls internally before showing you anything. I’d check your supervisor’s logging settings or add some custom logging to track what tools are actually firing at the supervisor level. The shortened outputs are probably because your supervisor is set to summarize responses for efficiency. Look for any response processing or truncation settings. I fixed similar truncation problems by tweaking the supervisor’s output formatting logic. Since LangSmith shows everything working fine, your core logic is solid - that’s actually good news.
Skip the supervisor config headaches - automation’s your answer here.
I’ve hit this same wall with multi-agent setups. Supervisors turn everything into black boxes where you can’t see what’s happening or control the flow properly.
Ditch the supervisor debugging and automate the whole thing instead. Chain your agents together in automated sequences while keeping full visibility at each step. You’ll get the same detailed outputs and tool transparency as running them solo.
Automation also lets you throw in custom logging and formatting wherever you want. All the coordination benefits, none of the visibility loss.
I’ve built these systems before - kills both the tool visibility and output truncation problems. Plus you control exactly how agents share data.
Latenode makes this workflow automation dead simple. Orchestrate your entire multi-agent sequence without losing control over inputs, outputs, or execution visibility.
yeah, i’ve run into this too - ur supervisor’s basically swallowing the outputs. set verbose=True on the supervisor agent and check for any middleware or response handlers that might be processing stuff. also, supervisors sometimes have their own token limits separate from individual agents, which can cause truncation.
Yeah, this is totally normal with supervisor architectures. The supervisor coordinates everything and handles outputs differently than when agents run directly. You’re seeing limited tool visibility because the supervisor bundles all the tool calls together before showing them - so you only get the final result instead of each individual call. For the cut-off outputs, your supervisor probably has built-in compression or summarization to avoid flooding the interface with tons of text from multiple agents. Check your supervisor config for output_mode or verbosity settings. Also look for max_tokens or response_length limits that might be chopping off your results. Since LangSmith shows everything’s working fine under the hood, it’s just how the supervisor displays info - not an actual execution issue.
We hit this exact problem 8 months ago. It’s not a config or settings issue.
Supervisors batch everything behind the scenes. Individual agents stream outputs directly to you, but supervisors collect that data first, then decide what to show based on their internal logic.
We fixed it by adding custom callback handlers to the supervisor itself. Hook into the tool execution events and force them to surface in your interface. LangChainCallbackHandler with verbose=True usually works.
For truncated outputs - supervisors often apply their own response formatting. Check for AgentExecutor wrappers around your supervisor. They compress multi-agent responses by default.
We created a custom supervisor class that inherits from the base one but overrides the output handling methods. Took 2 hours to implement and completely solved both problems.
Since LangSmith shows everything working, your agent logic is fine. This is just a display layer issue with how supervisors present information.
hey, sounds like ur supervisor might be holding back some output. have u tried toggling between invoke() and stream()? streaming often reveals all tool calls. also, check if return_intermediate_steps is set to false on the supervisor; that could be why ur seeing just summaries.