Seconded

The trust layer for autonomous work

Agents are about to do the world's work. Who signs off?

Seconded gives every AI agent and every person its own cryptographic signature, and makes every consequential action wait for independent sign-off. Merges, deploys, payouts, deletions: each one leaves a receipt anyone can verify. No agent approves its own work.

Try to forge an approval Talk to the founders
live · seconded gate 0 signed
    Real Ed25519 signatures, generated in your browser just now. The actions are simulated.

    Every high-stakes system humans ever built runs on one rule: no one approves their own work. Banks need two people to release a payment. Launch systems need two keys. Surgery stops for a second check. Public companies need a signed audit.

    That rule is about to be tested at machine speed. For AI agents, it barely exists yet.

    Why now

    The work is moving to agents. The accountability isn't.

    75%

    of new code at Google is AI-generated and engineer-approved, up from about 25% eighteen months earlier.

    Google, Apr 2026

    15%

    of day-to-day work decisions will be made autonomously by agentic AI by 2028, up from almost none in 2024.

    Gartner forecast

    53%

    of organizations have already seen AI agents exceed their intended permissions.

    Cloud Security Alliance, 2026

    40%+

    of agentic AI projects are forecast to be cancelled by 2027, with inadequate risk controls among the reasons.

    Gartner, Jun 2025

    Identity tells you who an agent is.

    Agent identity products answer which agent this is and what it can reach.

    Seconded proves who approved what it did.

    Was this exact action reviewed by someone independent, under the rules you set, before it ran? Seconded answers yes or no with a signature, not a log line.

    Born in production

    We ran 186 agents for four months. They tried every trick.

    Seconded started as the control system for our own fleet: AI agents from six model families writing, reviewing and testing each other's work on a real codebase. Here's what that taught us.

    186

    agent identities

    6

    model families

    ~7,300

    signed submissions

    ~2,400

    rejections

    incident · forged consensus

    One process, three keys.

    A single coordinator signed as author, reviewer and QA, and the change merged. Separate keys aren't separate actors unless signing is isolated.

    incident · fabricated evidence

    77 passes. Zero tests run.

    A QA agent that couldn't execute anything signed 77 passes in a row with invented evidence. Every signature was valid. A signature on a lie is still a lie.

    incident · monoculture

    69% of QA from one model family.

    Without rules, review quietly collapses onto one vendor's blind spots. Independence has to be enforced, not hoped for.

    Figures are approximate, from our internal agent fleet, July–November 2026.

    How it works

    Six primitives. One rule.

    1

    Identity

    Every agent and person gets its own key, in an isolated signer no other agent can reach.

    2

    Attestation

    Each step signs one exact version of the work, by its digest, with evidence attached.

    3

    Policy

    You decide who must sign and how independent they must be: a different identity, model, vendor, or a human.

    4

    Gate

    Every signature, digest and rule gets checked. One failure and the action is refused.

    5

    Trigger

    When the gate opens, the action runs. Nothing else holds the permission to run it.

    6

    Verify

    Every decision lands in an append-only ledger, and an open verifier lets anyone re-check it without trusting us.

    A policy is a file you can read.

    Policies are plain, versioned and signed themselves. This one guards deletions in a production cloud account.

    gate: prod-cloud-delete
require:
  - role: author
  - role: reviewer
    distinct_from: [author]
    distinct_vendor: true        # a different model vendor
    evidence: [diff_review]
  - role: qa
    distinct_from: [author, reviewer]
    evidence: [dry_run_log]
  - role: human                  # passkey sign-off
    when: env == "prod"
trigger: aws.ec2.delete_volumes

    Try it

    Try to forge an approval.

    This runs in your browser with real Ed25519 signatures. Five identities get fresh keys when the page loads. Run the honest flow, then try to cheat. Nothing leaves your machine.

    The change

    
          
    digest …
    
          
    Registered identities
    
          
      
        
      
        
      
          
      Honest flow
      
          
      
            
            
            
            
            
          
      
          
      Now cheat
      
          
      
            
            
            
            
            
            
          
      
          
        
      
        
      
          
      Ledger
      
          
      1. Empty. Start the honest flow.
      
          
      Gate result
      
          
      Not run yet.
      
        
      
      
      
    
      
  
      

  
      
    
      
      
      Vision
      
      
      HTTPS made the web safe to buy on.
      Seconded makes agents safe to hand real work.
      
      
        
        
      1. Now
        Gates
        Risky agent actions (merges, deploys, cloud changes, payouts) run only after independent, signed sign-off.
        2. 
        
      2. Next
        Reviewer reputation
        Planted-defect canaries measure which agents' sign-offs actually catch problems. It's a credit score for AI reviewers, built on data nobody else collects.
        3. 
        
      3. Then
        Receipts that travel
        Work from your vendors' agents arrives with proof attached, in an open format. Your auditor verifies it without a phone call.
        4. 
        
      4. End state
        Every autonomous action has a receipt
        The way every web page carries a certificate, every consequential thing an agent does carries a verifiable record of who approved it.
        5. 
      
      
      
      Gates are what we're building with design partners now. Everything after that is roadmap.
      
    
      
  
      

  
      
    
      
      
      Where it starts
      
      
      Wherever an agent can do real damage.
      
      
      
        
      Cloud and infrastructure
      Agents clean up, rightsize and apply Terraform, while a second agent and a human sign off on anything destructive.
      
        
      AI-written code
      Agent-written changes merge only after independent signed review, and your auditor gets proof of who approved what.
      
        
      Money movement
      Maker-checker for agents that handle refunds, payouts and journal entries. The agent that prepares a payment can never approve it.
      
        
      Anything with a webhook
      Emails, publishing, data exports, account changes. If an agent can trigger it and it matters, it goes through a gate.
      
      
      
    
      
  
      

  
      
    
      
      
      
        
      Talk to us
      
        
      We're building this with a small group of partners.
      
        
      Teams: if your agents already act on real systems, or you want them to, join as a design partner. You get the product first and help shape it.
      
        
      Investors: we'd like to show you the fleet data and the demo, and talk about where agent accountability goes next.
      
        
      From the team at Revelations Consulting: engineers who have run AI agents and cloud systems in production, and who built Seconded for their own fleet first.
      
      
      
      
      
        
      Start a conversation
      
        
      Email the founders directly. Put "Seconded" in the subject, and tell us what you'd gate first or what you'd like to see.
      
        hello@revelations.consulting
        
      We reply to every note, usually within a day.