Services / AI agent engineering
New serviceSoftware built by agents. Signed off by senior engineers.
Voyager is our in-house orchestration system. It runs AI coding agents on your backlog, from ticket to merged change, while our senior engineers set the standard and validate every piece of work.
ACME-142 · Linear · priority high
Add CSV export to the invoices list
- 09:12Claimednext eligible ticket, no blockers
- 09:13shipimplemented in an isolated worktree
- 09:17Verifiedtests, lint, type check and build pass
- 09:18Merge request !218 openedafter an independent review
- 09:21review:autofix2 review threads addressed, CI green
- 09:26Approvedby a senior WheelHouse engineer
- 09:26Mergedticket moved to Done
- faster development
- up to
- 10x
- lower cost for the same work
- up to
- 5x
- per client, on our premises
- one
- Mac mini
- validated by a senior engineer
- every
- change
Figures are upper bounds, not averages. Gains depend on how well specified the work is and the state of the codebase.
What Voyager is
Agents do the typing. Engineers do the engineering.
Voyager takes well-defined tickets from your backlog and delivers them as tested, reviewed merge requests. It is built on the principles our engineers already work to: small changes, verified before review, merged only when everything is green.
It does not replace engineering judgement. Senior WheelHouse engineers shape the work, answer the questions agents raise, and validate every change before it reaches your main branch. You get the pace of an agent fleet with the accountability of a senior team.
How it works
From ticket to merged change
Your team keeps working in Linear and your code host. Voyager fits around them, asks when something is unclear rather than guessing, and every step is visible as it happens.
- 01 · voyager
Select
Picks the next eligible ticket from Linear, by priority.
- 02 · agent · ship
Build
Implements it in an isolated worktree, with focused tests.
if anything is unclear
operator or clientAsk, don't guess
Unclear requirements become a question. Other work carries on.
- 03 · agent · ship
Review
Your checks and an independent review, then a merge request.
- 04 · voyager
Checks
Watches CI and reviewer feedback on every new commit.
until everything is green
review:autofixFix until green
Fixes what is justified, explains what is not, pushes again.
- 05 · engineer
Approve
Senior sign-off, then Voyager merges and closes the ticket.
Coming soon
Next: release management
Once changes merge, Voyager will take care of the release too: tagging it in git, publishing it on GitHub or GitLab, and keeping Linear's release pipeline up to date.
- coming soon
Tag release
Tags the merged work as a new version in git.
- coming soon
Publish release
Creates the release on GitHub or GitLab, with notes.
- coming soon
Track release
Keeps the Linear release pipeline in step.
Engineering principles
Built the way we build software
Speed only matters if the work holds up. These rules are part of Voyager itself, not guidelines an agent can choose to ignore.
Every change is a merge request
No direct pushes and no force-pushing. Each ticket gets its own branch, its own merge request and a clean history.
Verified before review
Agents run your repository's own tests, linters, type checks and builds before anyone is asked to look.
Green means green
A change merges only when required checks pass, required reviews finish and discussions are resolved, on the latest commit.
Bounded by design
Every run has limits on attempts, time and spend. Reaching one pauses the work with a clear reason, never an endless retry.
Questions, not guesses
Conflicting requirements or a genuine product decision pause the ticket for a person. Routine choices stay with the agent.
Measured, not assumed
Every run, intervention, cost and regression is recorded, so we can show you what was delivered and what it took.
Human oversight
Senior engineers stay in charge
Agents are fast. Judgement still comes from people who have built and run production software for years.
They shape the backlog. Tickets with testable acceptance criteria, sized and prioritised with you.
They answer the agents' questions. Decisions are made by people who know your product, not inferred by a model.
They validate every change. A senior engineer's approval is a required step before anything merges.
They own the architecture. Design, security and anything bigger than a ticket stays with engineers. Large tickets need their explicit go-ahead.
ACME-157 · paused in ship
Round per line, or once on the invoice total?
Acceptance criteria 1 and 2 disagree. AC1 rounds each line; AC2 expects the total to match the payment provider, which rounds once.
Your answer
- Round each line (AC1)
- Round once on the total (AC2)
- Something else
Meanwhile: ACME-158 in progress · ACME-159 in review
Security scans
Sensitive code gets a specialist security review
When a ticket touches a sensitive area, such as authentication, payments or personal data, Voyager has every version of its merge request scanned by security models that are not generally available.
Claude Mythos
Anthropic · trust programme
OpenAI Daybreak Blue
OpenAI · trust programme
WheelHouse has access to both through Anthropic's and OpenAI's trust programmes.
Two models, in parallel, read-only. They report what they find. They never change your code.
Every new commit is scanned again. A change merges only once the scan of its latest version passes.
Findings become review threads. The agent fixes them. A high or critical finding that is not fixed goes to a senior engineer.
Tested against known flaws. We evaluate the scan on code with planted vulnerabilities and clean controls, so we know what it catches and how often it raises a false alarm.
ACME-163 · voyager:security-scan
Save card details for repeat payments
scanners · read-only, in parallel
- highCard token written to the request logfound by both · fixed in 4e1c2a9
- mediumNo rate limit on adding a cardMythos · fixed in 4e1c2a9
- lowDecline reason shown in full to the userDaybreak Blue · explained, follow-up ACME-171
Security and isolation
Your own Mac mini. Your code stays yours.
Voyager runs on a fleet of Mac minis on WheelHouse premises. Each client gets a dedicated machine, so agents can work autonomously without sharing hardware, credentials or repositories with anyone else.
mini-01
another client
mini-02
another client
your-company
running ACME-142
mini-04
another client
mini-05
another client
mini-06
another client
your code host
GitHub or GitLab
your tracker
Linear
fleet portal
progress, questions, spend
Dedicated hardware
One Mac mini per client. Your repositories, credentials and data never share a machine with another client's.
Isolated work
Each ticket runs in its own worktree, scoped to the repositories and projects you approve.
Scoped access
Your machine holds only the access it needs for your code host and tracker, and nothing belonging to anyone else.
Central oversight
Our fleet portal shows every run's progress, questions and spend, so engineers see everything without pooling your code.
Fits your workflow
No new tools for your team
Developers and reviewers keep using Linear and your code host. Leave a comment on a merge request and Voyager picks it up. Agents run on leading coding models, chosen per step for the job.
- Linear
- GitHub
- GitLab
- CodeRabbit
- Claude Code
- Codex
Scheduled maintenance
Dependency upgrades on a cadence you choose, from weekly to annually. Voyager plans each cycle, raises the tickets and delivers them like any other work.
cadence: monthly · tickets: review first
Does AI write all of the code?
Agents write most of the code for the tickets Voyager takes on. Our senior engineers decide what gets built, answer the questions agents raise, and review and validate every change before it merges.
What kind of work suits Voyager?
Well-defined tickets with testable acceptance criteria: features, fixes, integrations and maintenance. Large or open-ended work starts with our engineers, who break it into tickets Voyager can deliver.
Where does my code live?
In your own repositories, and on the Mac mini dedicated to you at our premises while work is in progress. It is never placed on a machine shared with another client.
Which AI models does Voyager use?
Leading coding agents, currently Claude Code and Codex. We choose the model and effort for each step of the workflow, so you are not tied to a single provider.
Bring us your backlog
Tell us what you are building. We will show you what Voyager can take on, and where our engineers stay hands-on.
