TRUST INFRASTRUCTURE FOR AI AGENTS

Trust whatagents ship.

DuoAttest is the trust layer for agentic software delivery. It follows every coding task automatically, proves the exact task when a contract exists, and proves regression safety when it does not.

Install the Agent Skill See how trust is earned ↓
Independent verification · Real execution · Commit-bound evidence
INDEPENDENT OF THE CODING AGENTPROOF FROM REAL EXECUTIONBOUND TO THE EXACT COMMITCODE RUNS IN YOUR RUNNER
WHY TEAMS NEED A TRUST LAYER

Agents accelerate delivery. DuoAttest keeps confidence from becoming the bottleneck.

The product does not add another review queue. It turns each supported change into a decision a reviewer can understand and act on.

01LESS REVIEW WASTE

Review the risk, not every line.

See what the change had to prove, what actually ran, and where human judgment is still needed.

02FASTER SAFE DELIVERY

Move proven changes forward.

DuoAttest runs automatically with the pull request and returns a concrete merge decision instead of another wall of comments.

03INDEPENDENT BY DESIGN

The author never grades its own work.

The Agent can plan evidence. Independent Base/Head execution and deterministic policy decide the outcome.

04ADOPT WITHOUT MIGRATION

Use the CI and tests you already own.

Execution stays in your GitHub Runner. Setup is Agent-managed, and the Dashboard remains a read-only evidence center.

TRUST, BUILT INTO DELIVERY

One approval. Then trust becomes part of the delivery path.

The DuoAttest Skill gives your existing coding Agent the tools and operating rules it needs. Humans approve repository access once; the Agent manages setup and evidence, while DuoAttest stays independent.

01 / INSTALLGive SKILL.md to your Agent

Works with Codex, Claude Code, or any Agent that can run commands and call MCP.

02 / SAY“Connect this repository to DuoAttest.”

The Agent detects the repository, checks its status, and opens the only approval you need.

03 / APPROVEApprove the exact repository once

The Agent resumes automatically and handles setup, evidence readiness, and PR binding from then on.

YOUR AGENT OPERATESRepository setup → task contract → evidence readiness → PR binding
DUOATTEST DECIDESTASK VERIFIED / REGRESSION VERIFIED / BLOCKED / NEEDS ACTION
Start with the Skill
ONE CHANGE / ONE EVIDENCE CHAIN

Trust is earned by execution, not generated by a model.

This example shows the full product contract: the task claim, the same verification on Base and Head, and a decision bound to the exact commit.

payments / checkout-apiPR #184 · fix duplicate charge on retryHEAD 9f2c81a
01TASK CONTRACT

What this PR must prove

REQUIRED

A retried request creates only one charge

●
COMPAT

The normal payment path still works

●
COMPAT

Refund behavior does not change

●
DUOLA AI

Reads the task, diff, and existing tests; plans only the missing verification.

02DIFFERENTIAL EXECUTION

The same verification runs on both versions

BASE4b1a705
request_01charge.created
request_02charge.created
OBSERVED2 CHARGES
VS
HEAD9f2c81a
request_01charge.created
request_02duplicate.ignored
OBSERVED1 CHARGE
STABILITY REPLAY3 / 3
03MERGE DECISION
✓
TASK VERIFIED

Safe to merge

Duplicate charging is fixed. Payment and refund paths did not regress.

Required claims
3 / 3
New regressions
0
Certificate
9f2c81a
duola replay run_184
WHERE DELIVERY LOSES TRUST

The harder a change is to judge from the diff, the more the team needs independent evidence.

01
AI AGENT PR

An agent changed a dozen files at once

Turns the task into acceptance claims and maps each one to execution evidence. Reviewers do not have to infer what the code did.

CLAIMS → EVIDENCE COVERAGE
02
BUG FIX

The PR says fixed; the existing tests never covered it

Adds a disposable verification that never writes to the repo. Base reproduces the bug and Head removes it.

BASE FAIL → HEAD PASS
03
REFACTOR / UPGRADE

Behavior should stay the same, but the blast radius is unclear

Observes both versions under the same conditions. Only new, reproducible Head failures count as regressions.

BEHAVIOR EQUIVALENCE
THE CONTROL BOUNDARY

AI can plan. It cannot mark a PR green.

DuoLA AI understands the PR, locates impact, selects existing tests, and plans missing cases. Its output must pass structural validation. The final decision reads only real commands, assertions, observations, and stable replays.

See security and data flow
AI / NON-DETERMINISTICTask understandingImpact analysisVerification plan
RULES / DETERMINISTICSchema validationBase / Head executionEvidence policy
OUTPUTTASK VERIFIEDREGRESSION VERIFIEDBLOCKEDNEEDS ACTION
THE MISSING LAYER

The trust layer between code generation and production.

They are complementary. Teams can use Codex, Copilot, CodeRabbit, and DuoAttest together.

ProductDefault deliverableExecutionBase / HeadSupports a merge decision?
Codex ReviewDefect analysis and review commentsAssisted—Human judgment required
GitHub Copilot ReviewImprovement suggestions and comment review——Human judgment required
CodeRabbitIssues, walkthrough, and pre-merge checksPartial—Review-first
Traditional CIFixed job status and logsYesUsually noCovers configured commands
DuoAttestTask or Regression verdict, owner, next action, behavior evidence, and commit certificateUser runnerCore protocolYes; actionable owner and next step when blocked

Sources:CodexGitHub CopilotCodeRabbit

TRUST SHOULD RUN AUTOMATICALLY

Trust infrastructure should not become another workflow to maintain.

The operational surface is your coding Agent. The Dashboard is deliberately read-only: it shows decisions, behavioral witnesses, replay commands, and certificates without turning setup into another manual workflow.

Install the Agent Skill
01Agent owns operationsNO MANUAL SETUP
02Human owns consentEXACT REPOSITORY
03DuoAttest owns the verdictCOMMIT-BOUND PROOF