Forward-Deployed Engineering
A named senior engineer arrives with a team of agents, builds one workflow against your real records, cuts it over to real traffic and leaves. The agents stay. The system stays. Your people decide the exceptions.
The last mile
What we usually find
Forward deployment is the name for closing that distance from inside.
What goes forward
You meet the engineer before you sign and the same person stays until handover: senior, one engagement at a time, in your repository and your standups. The agents come with them: coding agents that write the connectors and the tests, evaluation agents that score every candidate, operations agents that watch the system once it runs. And the kit that makes leaving possible from day one.
On site for scoping and the first days of embedding; wherever your team runs after that.
The engineer
Senior. Named before you sign. One engagement at a time.
The agents
Build, evaluate, operate. They stay when the engineer leaves.
The kit
Connectors, an evaluation harness, a review desk, a runbook: patterns we have brought before, so nothing starts from scratch.
The first workflow
Every week ends in something you hold. The parts arrive in this order on every engagement, whatever the workflow.
One page: the workflow, the metric that will judge it, the person who owns it, the exit gate. Signed before anyone starts.
Access granted, repository cloned, the first pull request merged by day seven. That merge is how both sides know the access is real.
Real cases with known answers, agreed with the people who do the work. From here on, nothing is judged by a demonstration.
The connectors to your systems of record, written by the coding agents and reviewed by your reviewers, behind a feature flag.
The exception path: what the agents hand to a person, who that person is and what authority they have. Staffed before cut-over, not after.
Real traffic at real volume, rollback armed, monitoring live, the golden-set score as the argument for going ahead.
Your engineers release, roll back and retrain it with ours in the room and off the keyboard. The runbook is signed when that goes cleanly.
Current week: W0 Scope
After week two, every run is against your records.
The exception desk
Most of a workflow is routine, and routine is what the agents take. The rest goes to a desk: a few of your people with the context and the authority to decide, and an audit line for every case that reaches them. The desk is designed in week four and staffed before cut-over, because the exception path is where automation fails when it is an afterthought.
Three rules the agents work under
What reaches the desk is telemetry. A recurring exception becomes next month's automation.
The day we leave
Handover is the engagement's purpose, not its end. When your team has run the system without us, the engineer's access is revoked and what remains is yours: the agents, the code, the evaluation suite, the runbook, the decision record.
The agents
Running under your account, each with an identity, scoped permissions and an audit log.
The repository
Every commit yours from the first, reviewed by your reviewers, intellectual property assigned before week one.
The evaluation suite
The golden set and the harness, wired into your pipeline so a regression fails the build.
The runbook
Deploy, roll back, retrain, rotate, escalate; executed by your engineers before it is signed.
The context layer
How your organisation decides, written down for the next workflow and the next model, portable across vendors.
The owner
One person on your side who has released, rolled back and retrained it without us in the room.
Six weeks. One system on your floor. Your desk decides its exceptions.
Workflow after workflow, each with its own agents and its own exception path. The engineers gone; one desk for the department.
Department by department.
No one is replaced by a slide. Each step is a workflow in production, counted.
2027–2028
What the next two years change about this work, and what we already do about each.
Showing 2026
Today the engineer arrives with agents. Soon the agents write most of the integration and the engineer's job is judgement: what to automate, what counts as right, when to cut over. We already run it that way.
Code is generated. What is scarce is how your organisation decides, written down so any model can act on it. The context layer you keep is yours and moves with you between vendors.
Hours pay for approaching a result. We price a workflow against an agreed outcome, fixed before week one, and invoice at the gate, not the date.
Releases gated on evaluations, drift watched, model swaps regression-tested, cost per completed case known. Running agents is becoming a discipline with a pager. Your team inherits the instruments, or we run it until they do.
Scoped permissions, a named owner, an audit log: the same as a person with a badge. Regulators in several markets are moving to require it. We ship it now.
Your cloud, your premises, or no network at all. The oversight roles, the logs and the stop procedure that high-risk rules will demand are part of the handover, not an add-on.
The industry expects the human forward-deployed engineer to shrink to a sliver of each deployment within a few years. We design for it: what stays is your exception desk and your own people, able to run and extend what we built.
The next forward deployment is a robot fleet: policies fine-tuned on your floor, an engineer on site while the fleet learns, the same handover at the end. We deploy for physical work the same way.
Dates and rules move; the direction does not. Ask us what changed since this was written.
Two scales
Forward-deployed engineering means our engineers work with your teams, inside your systems and on your data, on site, remote or hybrid as the work needs, until a defined piece of work runs in production. It comes at two scales.
One function, one workflow or one department: agents do the defined work; your people decide the exceptions.
Task-specific agents, generative AI applications, knowledge systems and document workflows, each connected to the applications and data it needs. Each one ships with human approval where the work requires it, a log of every action and an evaluation against real cases. The offerings on this page are the ways in.
Find your starting pointDepartments, systems and data connected through governed agents, in phases.
A programme for organisations that want more than one solution: discovery across departments, integration with the systems they already run, then governed agents department by department, each phase evaluated before the next one starts. It is a direction we are actively developing, described as a programme, never as a finished result.
How the programme is phasedNeed one capability rather than an embedded team? Our services are scoped, built and handed over one at a time.
Where to start
Every engagement is the same system in a different configuration. Choose where you are; the shape changes to match and the offerings that fit are listed.
You know the work and the date it matters by. Six weeks from a signed scope to real traffic.
Six weeks, one named engineer
A system in production, its evaluation suite and its runbook
Runs inside every engagement
An evaluation suite, monitoring and the thresholds that page someone
The last two weeks of any engagement
A team that runs the system without us
It worked in the demonstration and never reached real users. We connect it, score it, harden it and cut it over.
Eight to twelve weeks
A pilot in production, with go-live criteria met and recorded
Scoped per system
Connectors, a data model and retrieval with access walls intact
Runs inside every engagement
An evaluation suite, monitoring and the thresholds that page someone
Several candidates and no scope. Two weeks inside your systems ends with one workflow, a metric and a fixed price.
Two weeks, one engineer
A scoped workflow, a success metric and a fixed price
Engagement Models
Reference
The terms before the first call
The Engineers
Reference
Who turns up on day one
Regulated data, your cloud or your premises, no egress. The oversight roles and the logs are built in from the first week.
Scoped per environment
A deployment with zero data egress and an audit trail
Scoped per system
Connectors, a data model and retrieval with access walls intact
Runs inside every engagement
An evaluation suite, monitoring and the thresholds that page someone
You want it built and kept running until your team takes it: a pod by the quarter, or operations by the month after handover.
Quarterly, two to three engineers
A shipping cadence your team keeps after the pod leaves
Monthly, after handover
Uptime, monitoring and iteration, reported monthly
The last two weeks of any engagement
A team that runs the system without us
Two weeks inside your systems; one workflow scoped in writing.
Two weeks, one engineer
A scoped workflow, a success metric and a fixed price
One workflow from prototype to production in six weeks.
Six weeks, one named engineer
A system in production, its evaluation suite and its runbook
A pod inside your repository and your standups, quarter by quarter.
Quarterly, two to three engineers
A shipping cadence your team keeps after the pod leaves
A stalled pilot hardened, integrated and put in front of real users.
Eight to twelve weeks
A pilot in production, with go-live criteria met and recorded
Your systems of record connected, your data modelled and governed.
Scoped per system
Connectors, a data model and retrieval with access walls intact
Inside your perimeter: your cloud, your premises or no network at all.
Scoped per environment
A deployment with zero data egress and an audit trail
A golden set, a harness and dashboards, so a regression is caught before a user is.
Runs inside every engagement
An evaluation suite, monitoring and the thresholds that page someone
Runbooks, decision records and a trained team, at a dated exit gate.
The last two weeks of any engagement
A team that runs the system without us
We keep running what we built, at a flat monthly fee, until you take it.
Monthly, after handover
Uptime, monitoring and iteration, reported monthly
Engagement Models
The shapes we work in, how long each takes and what we do not do.
Reference
The terms before the first call
The Engineers
Senior only, named before you sign, one engagement at a time.
Reference
Who turns up on day one
Every shape has a fixed fee agreed before week one and a written end.
TermsScope, timing, staffing and commercial terms are agreed per engagement, in writing.
Questions
Start
A thirty-minute call with an engineer. You leave knowing whether the workflow fits, what it would take and what it would cost.