I'm building an agent that calls tools, how do I record all the intermediate steps?

To record all intermediate steps in a tool-calling agent, the usual pattern is:

1. Log every model action

   • Prompt sent to the model
   • Model output
   • Any tool call name + arguments

2. Log every tool execution

   • Tool name
   • Inputs
   • Outputs
   • Errors / retries
   • Timestamps / latency

3. Store the full trace as a sequence of events A simple structure is:

[
  {"type": "llm_request", "prompt": "..."},
  {"type": "llm_response", "content": "...", "tool_call": {"name": "search", "args": {"q": "weather"}}},
  {"type": "tool_request", "name": "search", "args": {"q": "weather"}},
  {"type": "tool_response", "name": "search", "result": "..."},
  {"type": "llm_response", "content": "It is sunny."}
]

4. If you’re using an agent framework, enable tracing/callbacks Most frameworks expose hooks like:

   • on_llm_start, on_llm_end
   • on_tool_start, on_tool_end
   • on_chain_start, on_chain_end

5. Keep a run/session ID This lets you connect all intermediate steps from one agent invocation.

6. Persist to a datastore or observability system Common options:

   • SQLite/Postgres for simple persistence
   • JSONL files for easy debugging
   • OpenTelemetry / tracing backends for production
   • Framework-specific tracing tools if available

Example pseudo-code:

trace = []

def log(event):
    trace.append({**event, "ts": time.time()})

log({"type": "llm_request", "messages": messages})
resp = llm(messages)
log({"type": "llm_response", "output": resp})

if resp.tool_call:
    log({"type": "tool_request", "tool": resp.tool_call.name, "args": resp.tool_call.args})
    tool_result = tools[resp.tool_call.name](**resp.tool_call.args)
    log({"type": "tool_response", "result": tool_result})

If you want, I can also show:

• how to do this in LangChain
• how to do this in OpenAI tool calling
• or a framework-agnostic event schema for agent traces