Deposition
An open-source tool for recording and auditing the behavior of AI agents. It uses hash chains and Ed25519 signatures to ensure that agent behavior is tamper-evident and verifiable, and it supports mandate authorization constraints and causal traceability.
Timeline (last 90 days)
Deposition is an open-source tool that wraps agents with decorators to record every LLM call and tool call into a hash-chained jsonl file.
Each event hash in Deposition covers the previous event, so editing, deleting, or reordering events will be pinpointed by `depo verify` to the specific event that breaks the chain.
Deposition adds optional Ed25519 signatures to prevent attackers from rewriting the entire file and recomputing the hashes, upgrading the records from "tamper-evident" to "forgery-proof."
When initializing Deposition, a mandate can be passed in to declare constraints such as allowed tools, maximum cost, and maximum steps, and the mandate is sealed into the same trace.
Deposition's `depo audit` compares the run records against the mandate, points out out-of-bounds events, and links them to the model turn that caused the violation.
Every tool call in Deposition links back to the model message that requested it, supporting reverse tracing of "why this was done."
Deposition claims compatibility with LangGraph 1.2, CrewAI, OpenAI Agents SDK, Pydantic AI, and AutoGen, and the author tested against the real packages and placed probes on tool functions.
Deposition currently does not support capturing streaming responses: requests are recorded, and responses are marked as not captured.
Deposition's signing keys are stored on the same machine as the traces, making its security weaker than expected.
Deposition does not intend to replace LangSmith or Langfuse; the latter are monitoring tools, whereas Deposition targets the scenario where "someone later asks you to prove what the agent did."
3 more timeline events
All Facts (13)
Deposition is an open-source tool that wraps agents with decorators to record every LLM call and tool call into a hash-chained jsonl file.
50%UnverifiedEach event hash in Deposition covers the previous event, so editing, deleting, or reordering events will be pinpointed by `depo verify` to the specific event that breaks the chain.
50%UnverifiedDeposition adds optional Ed25519 signatures to prevent attackers from rewriting the entire file and recomputing the hashes, upgrading the records from "tamper-evident" to "forgery-proof."
50%UnverifiedWhen initializing Deposition, a mandate can be passed in to declare constraints such as allowed tools, maximum cost, and maximum steps, and the mandate is sealed into the same trace.
50%UnverifiedDeposition's `depo audit` compares the run records against the mandate, points out out-of-bounds events, and links them to the model turn that caused the violation.
50%UnverifiedEvery tool call in Deposition links back to the model message that requested it, supporting reverse tracing of "why this was done."
50%UnverifiedDeposition claims compatibility with LangGraph 1.2, CrewAI, OpenAI Agents SDK, Pydantic AI, and AutoGen, and the author tested against the real packages and placed probes on tool functions.
50%UnverifiedDeposition currently does not support capturing streaming responses: requests are recorded, and responses are marked as not captured.
50%UnverifiedDeposition's signing keys are stored on the same machine as the traces, making its security weaker than expected.
50%UnverifiedDeposition does not intend to replace LangSmith or Langfuse; the latter are monitoring tools, whereas Deposition targets the scenario where "someone later asks you to prove what the agent did."
50%UnverifiedDeposition is licensed under Apache 2.0, has no required dependencies, and most of its 14 examples can run without an API key.
50%UnverifiedIn Deposition, the mandate is fixed and sealed at initialization as the baseline for subsequent audits, unlike runtime dynamic policies.
50%UnverifiedDeposition binds "what the rules are" and "what happened" into the same evidence chain, eliminating after-the-fact verbal disputes about the rules.
50%