AgentLedger Adapters

One-liner audit trails and policy enforcement for the most popular Python agent frameworks.

from agent_ledger_adapters import build_ledger
from agent_ledger_adapters.langchain import AgentLedgerCallbackHandler

ledger = build_ledger("agent_ledger.db", policy_path="policy.json")
chain.run(question, callbacks=[AgentLedgerCallbackHandler(ledger, agent_id="my-agent", enforce=True)])

Built on AgentLedger (vendored, zero dependencies). Every agent action lands in a tamper-proof SHA-256 chain; with enforce=True, policy violations raise and stop the offending action.

Supported frameworks

No framework is required to install this package — imports are lazy, so only the adapter you use pulls in its framework.

What gets logged

chain_start/end, agent_action, tool_call, tool_result, tool_error, llm_start/end/error, agent_start/end, run_start/end, handoff — with run IDs, inputs, outputs, and error messages (truncated).

Policy enforcement

{
  "version": 1,
  "default": "deny",
  "policies": [
    {"agent": "my-agent", "action": "delete_*", "target": "*", "decision": "deny"},
    {"agent": "my-agent", "action": "deploy", "target": "production", "decision": "require_approval"}
  ]
}

• deny → PolicyViolation raised (with errors="raise"), violation logged
• require_approval → logged with policy_status: pending_approval
• errors="ignore" (default) → auditing never breaks your agent

Run the tests

cd tests && python3 test_adapters.py   # mock framework objects, no frameworks needed

MIT