Autonomous Backlog Engine · Private beta
Run ABE on your backlog.
ABE turns tickets into verified pull requests. It reproduces the bug, writes the smallest fix, tries to break it, proves it, and applies your policy before an engineer spends a minute on it.
Works with Jira, Azure DevOps, GitHub, and Bitbucket. No model keys to manage. Humans keep the merge button.
How it works
Six stages. Two of them argue.
Every ticket moves through the same pipeline. Four stages are deterministic platform work. Two are AI agents with opposed jobs.
Scout
Reads the ticket, finds the owning code, checks for duplicates and fixes that already exist.
Repro
Reproduces the failure and encodes it as a failing test. Graded R0 to R3 so you know how strong the proof is.
Maker
Writes the smallest fix that turns the test green, plus the tests that keep it green.
Breaker
An independent reviewer on a different model family. Its only job is to break the fix.
Proof
Runs the build and the test suite and records the evidence. A narrative never counts as a pass.
Gate
Applies your policy to the evidence. Ship, review, or stop. Fails closed.
AI agents. Everything else is code you can read.
Separation of duties
Every fix has a Maker and a Breaker.
One model grading its own work is not a review. Two models with opposed jobs is.
Maker
Finds the root cause, writes the fix and its tests, and drafts the pull request with every claim tied to evidence.
It cannot widen scope, weaken a test to get a pass, or add a dependency without policy saying so.
Breaker
Reviews the change with fresh eyes and one question: what would have to be true for this fix to be wrong?
Every finding is graded. The Maker can answer a finding with evidence. It cannot delete one.
Two models. Opposite jobs. Findings that can be answered but never erased.
Evidence, not assertions
A pull request you can audit in five minutes.
The reviewer never reconstructs a run from raw logs. Every ABE pull request carries its own proof.
- The reproduction and how strong it is
- The failing test, then the passing run
- Every Breaker finding and how it was resolved
- The proof ledger: build, tests, code health
- The Gate decision and the policy version it used
- Which models did the work, and what it cost
| Reproduction | R3 encoded as a failing test |
| Test run | red to green · 148 passed · 0 skipped |
| Breaker findings | 1 · S1 total could go negative when the discount exceeds the subtotal · resolved |
| Proof ledger | build pass · tests pass · code health pass |
| Gate | low risk · policy v3 · review-ready |
| Models | Maker Anthropic · Breaker OpenAI |
Control
Your engineers stay in charge.
ABE asks for trust one rung at a time and shows its work at every rung.
Autonomy
Dialed up, never assumed.
Start with review-ready pull requests. Auto-merge is earned per tier from measured outcomes and revoked automatically when they slip.
Models
Pick the model. We hold the keys.
ABE runs on Amazon Bedrock and other providers under accounts we manage. You choose which model sits in which seat from settings. Nothing to provision, no keys to rotate. Enterprise deployments can run inside your own AWS account.
Audit
Every run leaves a trail.
A complete record of every run: what ABE did, what it found, and why the gate decided. Spend ceilings per day and per ticket. Nothing merges on a narrative.
Replay
Before you trust it, replay it.
Point ABE at a set of tickets your team already resolved. It works them without seeing the accepted fix. You compare its pull requests and verdicts with what actually shipped, before a single live ticket.
- 01Pick the set. Resolved tickets from the last quarter, sensitive paths excluded.
- 02ABE works them blind. Same pipeline, same Breaker, no access to the accepted solution.
- 03Compare. Acceptance rate, Breaker catches, review minutes, and cost per accepted change.
- 04Decide the rung. The replay tells you where on the ladder to start.
Status
Private beta now. Design partners open Q4 2026.
ABE is in production with a private beta customer. A handful of design-partner seats open this quarter. A pilot starts with a replay of your own tickets.