Agent flight recorder for lending agents
Tamper-evident logs of what lending agents did, with cheap monitors that flag consequential decisions and link the evidence, formatted for bank model-risk reviewers, loan QC and independent reviewers.
Competitive
Written for: Small startup team
Next step
Call Vesta's head of compliance or partnerships. Ask what evidence their lender customers' model-risk and QC teams demand before they let agent swarms touch a loan file.
Why now
Promoted from watch. Agents are now in live lending workflows (Vesta's $30M round for mortgage agent swarms, Nivo's Lender Collect). MAS now requires independent review before AI goes live. Monitoring that runs on every step is affordable with Haiku 5.5 pricing, and probe-based monitors are an option for teams that run open-weight models.
What became possible
Agent monitoring got cheap and measurable. Goodfire's monitors read the model's internals and call a second model only when something looks wrong. A new probe paper reaches 98.8% AUC at detecting sabotage, beating a text-reading monitor. AgentMonBench provides a benchmark for flagging consequential decisions and finding the evidence behind them.
Two ways in
- Loan-file agent recorder
An SDK that records every agent action on a loan file, hash-chains the records so they cannot be altered unnoticed, scores each action for consequence and reversibility, and exports a review pack matching SR 11-7 and MAS expectations.
- Customer
- Lending-agent vendors such as Vesta and Nivo, who must satisfy their lender customers' risk teams
- Model
- Usage-based per recorded agent run, plus a paid tier for exporting audit packs
- Real-time intervention hook
A streaming monitor that pauses an agent before irreversible actions, such as sending a conditional approval or moving funds, when its path drifts from expected patterns.
- Customer
- Fintechs and lenders running agents in production
- Model
- Per-seat platform fee plus monitored-step volume
Competition
Tamper-evident signed tool-call logs are already common in open source (Aileron, Signet, ActionProof, mcp-audit-gateway), and reversibility classification exists too (the search summary attributes it to an open-source proxy called Sagaz; irredux is on PyPI), so neither sets a product apart. Bank agent-governance platforms (Prefactor, Arthur, Arize, FairPlay, Space and Time) and AI-insurance players (AIUC, Armilla, Klaimee, Mount, Testudo) have both ends of the market covered.
Full competition notes (16 competitors)
Tamper-evident signed tool-call logs are already common in open source (Aileron, Signet, ActionProof, mcp-audit-gateway), and reversibility classification exists too (the search summary attributes it to an open-source proxy called Sagaz; irredux is on PyPI), so neither sets a product apart. Bank agent-governance platforms (Prefactor, Arthur, Arize, FairPlay, Space and Time) and AI-insurance players (AIUC, Armilla, Klaimee, Mount, Testudo) have both ends of the market covered. The remaining gap is the layer in between: (1) a mapping layer that keeps up with the rules. Secondary sources say an April 2026 OCC bulletin (2026-13, not checked against the primary source) took generative and agentic AI out of SR 11-7's scope, so banks need a crosswalk to other authorities, not to SR 11-7 itself. (2) Turning the runtime record of what an agent actually did into ongoing evidence for AIUC-1 controls (E015 logging) and MGA questionnaires, since insurers currently judge risk from a snapshot. (3) Containment proofs (the agent can't widen its own limits) presented as examiner-ready findings, which no open-source tool offers. A new entrant should reuse the open-source signing layers, partner early with an MGA such as Armilla or Klaimee, and pick a name other than AgentLedger. The main risk is that AIUC or the governance platforms add this feature themselves.
- Aileron (open source)
Calls itself a 'flight recorder for AI agents'. Hash-chained tool-call journal with signed checkpoints, policy enforcement, incident reports and offline verification. Very close to the logging core of the idea.
- Signet (prismer-ai)
Gives each agent an Ed25519 identity, signs every MCP tool call and keeps a hash-chained receipt log that can be verified offline.
- mcp-audit-gateway
MCP gateway that writes every tool call to a signed, hash-chained, append-only audit trail.
- ActionProof
Ed25519-signed, tamper-evident receipts for each agent action, recording what was done, by which agent, when and on whose authority.
- agent-audit-trail
HMAC-signed, tamper-evident log entry for each agent action, with no infrastructure needed.
- Sagaz / irredux (reversibility classification)
The search summary describes Sagaz as an open-source MCP proxy that labels agent effects Reversible, Compensable or Irreversible before they run, with preview and checkpoints. Linked page is irredux on PyPI; I did not open it to confirm it is the same project. Covers the consequence and reversibility scoring feature.
- Prefactor
Agent governance platform with a dedicated agent-audit use case for banking. Direct competitor for the bank buyer.
- Space and Time + W3.io
Verifiable, tamper-resistant execution and audit trail for AI-driven financial workflows. Reported at 200k+ workflows a day in production. Publishes on agents and MRM gaps.
- Arthur AI
AI governance and monitoring vendor that markets agent MRM and compliance tools to banks.
- Arize (agent governance for banking)
Observability vendor that positions agent tracing as action-level evidence for bank MRM.
- FairPlay AI
Publishes an MRM manual for agentic systems and sells to bank model-risk teams.
- AIUC (AIUC-1 standard + insurance)
Raised a $15M seed. Runs AIUC-1, the agent certification standard whose E015 control specifies the logs it expects. Markets agent coverage up to $50M. Sets the evidence format insurers want.
- Armilla
AI liability MGA backed by Lloyd's capacity. Markets cover for autonomous-agent decisions, tool use and escalation failures.
- Klaimee (YC)
Liability insurance built for AI agents. Underwrites from a public-data scan, a 30-question governance assessment and 100+ adversarial probes. Raised $5.5M.
- Mount (YC)
Building an AI-agent insurance carrier. Scans and red-teams deployments, puts a number on operational risk and insures what's left.
- agentic-ledger ('AgentLedger')
Open-source agent observability proxy that already uses the AgentLedger name. Name clash.
Checked 2026-10-07
Risks
- White-box probes need open weights. Teams on closed APIs only get text monitors.
- Observability incumbents and agent platforms (Google, AWS) may add audit logs themselves.
- Risk teams may not yet know what format they want, so sales cycles could stretch.
Strong team fit; open space is the weak point (5).
tech 7 · pull 7 · space 5 · fit 8 · build 7
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
- Goodfire says its new ‘inside-out’ monitors catch rogue AI agents at a fraction of the costTechCrunch AI
- Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized DeceptionarXiv
- What Did the Agent Actually Do? Evidence-Grounded Oversight for Long-Horizon AgentsHF Daily Papers
- OnTrack: Real-Time Monitoring and Intervention in LLM Agent Trajectories via Streaming Structure-Aware Optimal TransportarXiv
- Vesta raises $30M to bring swarms of agents to mortgage lendersTechCrunch Venture
- Nivo launches AI assistant for lendingFinextra
- Singapore central bank issues risk rules on banks' AI useFinextra
What happened next
- +7d · 14 Oct 2026 nothing yet. No launches, funding rounds, or acquisitions of lending agent flight recorders or related auditing solutions were found in the week after October 7, 2026. While broader agent governance and audit trail discussions appear in financial-services vendor content, no specific product launch or company event in this niche occurred in that timeframe.
Status history
- 2026-10-07 Added to the watchlist
- 2026-10-09 Promoted from the watchlist: Today's evidence made it buildable
- 2026-10-09 Flagged as buildable