What is an agentic delivery pipeline?

A governed workflow where AI agents write plans, code and tests, a different agent reviews them, and people approve plans, review changes and own merges and releases, with evidence recorded at each step.

Do we have to use your tools?

No. The workflow runs in your repositories and CI, with Claude Code and Codex or the agents you already use, and the skills are files you own.

Issues in, reviewed releases out: agents build, people approve

An agent delivery pipeline turns an approved issue into a reviewed pull request. An agent writes a brief and a plan, an engineer approves them, the agent builds and tests in an isolated branch, a second agent from a different provider reviews for spec, standards and security, and a person reviews and merges. Agents never approve, merge or deploy, and every step leaves evidence in the repository.

Book a demo How an engagement runs

Who owns the mistake? The code owner who reviews and the engineer who merges. Agents can open pull requests, never merge them.

Built for VP engineering · Engineering manager · Head of platform · CTO

acme / billing / pull / 2481Plan approved by AshaIllustrative run

fix: round GST per line before totalling the quote

#2481 · from LIN-142 · branch agent/lin-142-gst-rounding

 export function quoteTotal(lines: QuoteLine[]) {
−  const subtotal = sum(lines.map((l) => l.qty * l.rate));
−  return round(subtotal * (1 + GST_RATE));
+  // GST is rounded per line, as on the tax invoice.
+  return sum(lines.map((l) => round(l.qty * l.rate * (1 + l.gstRate))));
 }

Types and lint
Tests: 214 passed, 6 new
Secret and dependency scan
Review by Codex: clean after 1 round
Code owner review: 0 of 1

Opened by agent, via Priya+2 −2 · 1 fileMerge: people only

The work

How it runs today, and what the agent does instead.

Today

01Coding is faster; review is now the bottleneck.
02Pull requests get bigger and wait longer for a person.
03Nothing guarantees that a plan was agreed or that tests are real.
04Decisions and reasons disappear with the chat session.

With the agent

01BriefReads the code and current docs, asks what the issue doesn't say, and writes a brief for approval.
02PlanOffers two or three approaches, names the files and interfaces, and starts with a failing test.
03BuildImplements in an isolated branch with short-lived, least-privilege credentials and no production secrets.
04VerifyRuns the full suite and scans fresh, and records the evidence against the branch head.
05ReviewA second agent from another provider reviews spec, standards and security, with up to three fix rounds.

People decide

Where a person decides. Nothing leaves the building without them.

What comes out

Reviewed pull requests with tests and evidence
A changelog and a plain-language record of what changed and why
Weekly delivery metrics against the baseline

EngineerThe brief, then the plan, before any code is written.
Code ownerRequired human review on every pull request.
Release ownerMerges and releases. Agents can't.

The number

We agree the number before we build. Then we report against it every week.

Systems it works in

GitHub
GitLab
Jira
Slack
Sentry
Linear
Claude Code
Codex
Your CI

What we baseline and report on for issues in, reviewed releases out
Metric	How it's measured
Lead time for changes	Issue approved to change in production.
Change failure rate	Held at or below baseline.
Review rounds	Fix rounds per pull request before merge.
Merge rate	Agent pull requests merged without a rewrite.

Market context

What others are seeing. Their numbers, with their sources. Not ours.

We publish our own numbers for a client only when the client agrees. These are here for context.

<10%
rise in team output when coding gets 30–40% faster but planning, testing and release stay manual.
Forrester, 2026(opens in a new tab)
DORA 2025
AI adoption now correlates with higher delivery throughput and still with higher instability. The system around the tools decides the outcome.
DORA State of AI-assisted Software Development 2025(opens in a new tab)

Fair questions

What people ask before they trust it.

AI-written code is insecure.: It passes the same gates as human code, plus a second-agent review and dependency and secret scanning. Agents run sandboxed and never see production secrets.
Our developers will resist.: Developers approve every plan and review every change. We start with the work nobody wants: tests, upgrades, flaky-test fixes and small backlog items.

Related work

Where we've built this.

VegaStack, on client and internal workSoftware · Aug 2026 onwards
Our own delivery workflow: issues in, reviewed releases out
Agents build, a different agent reviews, people approve and ship. Every step leaves evidence.
In daily useRead

Questions

Issues in, reviewed releases out: the short answers.

Also on this desk

Coding-agent rollout and governance Upgrade and migration sprints For software and technology

Send us ten real documents. We'll show you what an agent does with them.

A one-hour call with a founder and an engineer. Bring a BOQ, a stack of POs, last month's invoices or a backlog, and leave knowing what would work, what wouldn't, and what it would take.

Book a demo or email founders@vegastack.com

170+ projects · 65 companies · founders on every engagement

Issues in, reviewed releases out: agents build, people approve

Who owns the mistake? The code owner who reviews and the engineer who merges. Agents can open pull requests, never merge them.

Built for VP engineering · Engineering manager · Head of platform · CTO

export function quoteTotal(lines: QuoteLine[]) {

const subtotal = sum(lines.map((l) => l.qty * l.rate));

return round(subtotal * (1 + GST_RATE));

// GST is rounded per line, as on the tax invoice.

return sum(lines.map((l) => round(l.qty * l.rate * (1 + l.gstRate))));

}

Metric

How it's measured

Lead time for changes

Issue approved to change in production.

Change failure rate

Held at or below baseline.

Review rounds

Fix rounds per pull request before merge.

Merge rate

Agent pull requests merged without a rewrite.

Issues in, reviewed releases out: agents build, people approve

How it runs today, and what the agent does instead.

Where a person decides. Nothing leaves the building without them.

We agree the number before we build. Then we report against it every week.

What others are seeing. Their numbers, with their sources. Not ours.

What people ask before they trust it.

Where we've built this.

Our own delivery workflow: issues in, reviewed releases out

Issues in, reviewed releases out: the short answers.

What is an agentic delivery pipeline?A governed workflow where AI agents write plans, code and tests, a different agent reviews them, and people approve plans, review changes and own merges and releases, with evidence recorded at each step.

Do we have to use your tools?No.

Send us ten real documents. We'll show you what an agent does with them.

Issues in, reviewed releases out: agents build, people approve

How it runs today, and what the agent does instead.

Where a person decides. Nothing leaves the building without them.

We agree the number before we build. Then we report against it every week.

What others are seeing. Their numbers, with their sources. Not ours.

What people ask before they trust it.

Where we've built this.

Our own delivery workflow: issues in, reviewed releases out

Issues in, reviewed releases out: the short answers.

What is an agentic delivery pipeline?A governed workflow where AI agents write plans, code and tests, a different agent reviews them, and people approve plans, review changes and own merges and releases, with evidence recorded at each step.

Do we have to use your tools?No.

Send us ten real documents. We'll show you what an agent does with them.