Autonomous software delivery

From request to main.
Autonomous, with evidence.

Cellvoy takes a Jira epic from a raw request all the way to merged code on main. Every rule that matters is enforced in code, every claim is backed by a real run, and an independent QA Agent decides what ships.

See how it works
  1. Request
  2. Epic
  3. Tickets
  4. QA
  5. main

The problem

Coding agents write fast. Delivery still runs on people.

Claims, not evidence

The agent says the tests passed. Nobody ran them where it counts.

Collisions at scale

Parallel agents touch the same files. Branches drift and conflict.

A manual last mile

Review, integration and the call to merge into main still eat senior engineers’ days.

The insight

The bottleneck isn’t the model. It’s the missing process.

Trustworthy autonomy = rules enforced in code + evidence from real runs + independent review.

What Cellvoy is

Four agents. One code layer that holds the keys.

Analyst

BA Agent

Turns a raw request into a PRD, an epic and tickets.

Writer

Implementer

Writes each ticket in its own isolated worktree.

Checker

Reviewer

A different, read-only model checks every acceptance criterion.

Gatekeeper

QA Agent

Decides whether the finished epic is fit for main.

Code layer: Git, validation and delivery. The only part of Cellvoy that commits, pushes, opens pull requests and merges.

End to end

From a request to main, in one flow

  1. Request

    A stakeholder need, in plain words.

  2. BA Agent

    PRD, Jira epic, tickets and the epic branch.

  3. Ticket loop

    Plan, reserve, implement, review, validate, merge into the epic.

  4. QA Agent

    One verdict on the whole epic.

  5. main

    Merged, with the evidence attached.

Human decision 1: approve the ticket list, then press Start.
Human decision 2: sign off only when QA asks for it.

BA Agent

From a vague request to a ready-to-run epic

  • Reads before it asks. Every repo in the project plus the chosen Confluence spaces.
  • Asks for decisions, not forms. Findings come with sources; the analyst picks.
  • Pushes back with evidence. Conflicts between the request, the docs and the code are shown, never silently resolved.

Ticket loop

Run the whole epic in parallel, without collisions

  1. Plan
  2. Reserve files
  3. Implement
  4. Review
  5. Validate
  6. Deliver

Isolated worktrees

Each ticket gets its own Git worktree and branch, cut from the epic branch.

File reservations

Two tickets never write the same file at once. Overflow is caught in code.

Capacity and resume

Concurrency limits per host and project. Sessions resume from their last stage.

Separate roles

The writer never grades its own work

planner model

Manager

Plans the ticket, names the files to reserve, runs the rounds.

developer model

Implementer

Edits only the reserved files on the prepared ticket branch.

a different model, enforced

Reviewer

Read-only. Sees the acceptance criteria and the real diff, not the implementer’s story. Returns MET or MISSING per criterion, with evidence.

The outcome contract is strict and fail-closed: malformed or inconsistent output never passes.

Evidence

Proof comes from real runs, not from the agent

Validation
Your repo’s own build and tests, run in its devcontainer.
Red run
Fed back to the same session; after 3 attempts the ticket stops with its reason.
Read-only
Validation can’t change the repo; git state is checked before and after.
Audit trail
Every prompt, output, result and decision is stored as an artifact.
Pull requests
Descriptions are built from those artifacts, not from a model’s story.

“Tests passed” is never taken on trust. Before delivery the code layer re-checks the PASS outcome and green validation.

Safety

Guardrails live in code, not in prompts

Agents can’t touch history

No commit, push, merge or rebase. HEAD is compared before and after every agent call.

Protected branches refused

Work on main, develop, release or hotfix branches is rejected outright.

One delivery path

Commit, push, pull request and merge happen in one resumable, idempotent step.

No secret leakage

Tokens and credentialed Git URLs never reach logs. Secrets are stored encrypted.

New · QA Agent

The last gate before main

Looks at

The epic and its blast radius

Everything the epic changed, plus the code that calls it and that code’s tests.

Against

Intent and fresh evidence

Acceptance criteria, the PRD, and validation re-run on the tip of the epic branch.

Produces

Blocking vs. backlog

Blockers become follow-up tickets on the same epic. Old debt goes to the backlog and never holds the merge.

Missing tests become findings, not QA-written code. Every epic leaves tests behind, so your regression suite grows with each one.

How QA decides

  • Blocking findings, rounds leftFollow-up tickets open on the same epic; the loop fixes them and QA runs again.
  • Blocking findings after 3 roundsHanded to a human to approve or reject.
  • Clean, touches sensitive pathsWaits for human sign-off; who approved, and when, is recorded.
  • CleanMerged into main automatically.

A “clean” verdict never merges over red validation. Sensitive paths (auth, payments and so on) are set per project.

Onboarding

Getting a repo ready is one pull request

  1. 1

    Inspect

    Cellvoy reads the repo and works out how it builds and tests.

  2. 2

    Generate and prove

    Adds the validation recipe, devcontainer and agent instructions, then proves they run.

  3. 3

    Your team merges

    Delivered as a pull request to review. Cellvoy never merges it itself.

“We onboard working repos. We don’t repair broken ones.”

Integrations

Fits the stack you already run

Work tracking

Jira · Confluence

ClickUp as a read-only task source.

Code hosting

GitHub · GitLab · Bitbucket

Pull and merge requests.

Feedback

Back in Jira

Status transitions, results and stop reasons on the issue.

Models

Bring your own

OpenAI, Anthropic, DeepSeek, Kimi, GLM and more. A model per role, chosen per project, with cost tracking per run.

Deployment

SaaS or on-prem

Managed by us, or run inside your own infrastructure.

Positioning

A delivery platform, not a coding assistant

DimensionIDE assistantsSingle-agent AI devsCellvoy
Unit of workLines and filesOne ticketThe whole epic, to main
Runs onA developer’s laptopThe vendor’s cloudSaaS or your infrastructure
EvidenceThe agent’s wordThe agent’s wordReal validation runs
ReviewA humanThe same agentSeparate reviewer + QA Agent
Parallel workNoneLimitedCapacity + file reservations
Git controlIn the agent’s handsIn the agent’s handsCode layer only
Merge to mainA humanA humanQA Agent, human on sensitive code

Proof

Cellvoy builds Cellvoy.

The QA Agent itself was delivered by Cellvoy’s own ticket loop: every backend and admin ticket ran in its own worktree and merged into the epic branch without a human touching Git.

Let’s talk about your backlog

Tell us about your engineering workflow. We’ll pick one real epic and run it end to end with your team watching the evidence.