When your agent fails at 2am, there's no log, no replay, no trace. Agent Basin captures every tool call, decision branch, and output — so you always know exactly what happened.
# Agent Basin — run captured 2025-11-12T02:17:43Z run_id: "a3f8c-29d1-..." agent: "customer-onboarding-v2" framework: "langchain" steps: - tool_call("search_crm", query="acme corp") ✓ 124ms - decision: account_found=true → branch "existing_customer" - tool_call("send_email", template="onboard_v3") ✗ FAILED - retry(1/3) ✗ FAILED # root cause found here error: "SMTP_AUTH_INVALID — credential rotated 6h ago" duration: 4.2s | tokens: 1,847 | cost: $0.0034
Model observability tools track tokens and latency. But agents are different — they branch, loop, call tools, and make decisions. When they fail in production, there's nothing to look at.
Standard logging misses most of what agents do. Tool calls, LLM prompts, branch decisions — all invisible.
You can't re-run the exact same inputs in the same state. The failure is gone, and so is any chance of understanding it.
Every agent run vanishes into the void. You can't query "show me all runs that hit the payment tool and failed."
Agent Basin wraps your existing agent framework with a lightweight SDK. One import. Zero agent changes required.
Every tool call, LLM prompt, decision branch, and output captured with microsecond timestamps and full context.
Replay any run exactly as it happened — pause, rewind, inspect state at each decision point.
Query runs by tool, error, duration, cost, or any custom attribute. Find the exact run you need in seconds.
Get notified instantly when a run fails, loops unexpectedly, exceeds cost thresholds, or hits a new error pattern.
Compare two runs side-by-side to see exactly where their decision trees diverged.
Per-run token usage and cost breakdown by step, tool, and model — with trend alerts.
Agent Basin works with LangChain, AutoGPT, CrewAI, custom frameworks — anything that runs in Python.
pip install agentbasin
basin.capture(agent)
Runs stream in live. Search, replay, and set alerts immediately.
import agentbasin as basin from langchain.agents import AgentExecutor # Your existing agent — unchanged agent = AgentExecutor(...) # Wrap it — that's it agent = basin.capture( agent, project="my-prod-agent", api_key="basin_..." ) # Every run now captured automatically result = agent.run("Process refund for order #8821")
No credit card required. Free plan includes 1M events per month — enough for most early-stage teams.
View pricing →Stop flying blind. Connect your first agent in under 5 minutes — free.
Connect your first agent — free