Four engagements, one alternative entry, and a module catalogue.
Every engagement is fixed-scope and fixed-fee against a bounded outcome. Every one names the
rung it covers and the rungs it does not. Nothing here is sold by the hour, and
nothing is sold as a retainer before there is something in production worth operating.
Quoted as a sequence
Entry plus core is $105k–$190k. We say that in the first meeting.
Discovering it in month three reads as a bait-and-switch, and it is.
Outcome, never hours
Fixed fee per bounded outcome. New priorities replace old ones or take a change order.
Our efficiency is our margin problem, not your discount.
Every engagement is exit-ready
Artifacts live in your repositories and your CI. The acceptance test is your operator
running the documented procedure without us in the room.
Rung 1 · The door
Agent Runtime & Control Baseline
Three to four weeks, fixed fee, working range $30k–$50k. Read-only first,
so it clears security review without a fight. It is designed to be worth the money even if
you never buy anything else — and roughly a third of its value is discovering that a thing
you believed was true about your estate is not.
What we do
Estate inventory. Every AI and agent workload already running, its owner, its trigger, its blast radius, and its cost — including the ones nobody told you about
Tool plane and identity assessment. What agents can call, under whose credentials, with what scope, and what happens on a bad day
Lifecycle and containment audit. Including a live disable-survives-restart test on a non-production agent
Outcome ledger instrumentation. Flow, agent, adoption, and business tiers wired to real telemetry — before anything is built
Golden task suite. 20–40 representative engineering tasks from your team’s actual recent work, with known-good outcomes and deliberate false-completion traps
Data path review. Which datasets agents already read or write, classification gaps, and whether a masked test-data path exists at all
What you keep
A ranked gap report with a costed 90-day plan, prioritised by blast radius rather than by what we sell
The golden task suite, in your repository, runnable in your CI
The outcome ledger definitions and the baseline numbers
The agent registry as it stands, with named owners
Evidence of what the containment test actually did
When this is not right for you: if you have no platform —
no CI/CD, no IaC, no cloud governance — there is nothing to install a runtime onto and
the honest sequence is platform first. We will say so.
Rung 2 · What you came for
Runtime, Tool Plane & Guardrails
Six to eight weeks, working range $75k–$140k. This is the engagement a
platform lead can describe to their own leadership in one sentence: agents can now do work
here, inside limits that hold, and we can prove the limits hold.
Built
Environment fabric — ephemeral, isolated, fast environments agents can create, break, and destroy, with seeded and masked data. Provisioning time is a tracked SLO, because an agent waiting 25 minutes cannot iterate
Tool and action plane — the permissioned catalogue of what agents may do. Each tool a versioned contract with typed I/O, idempotency semantics, prohibited operations, a dry-run mode, and trace correlation
Identity and continuous authorization — per-agent, per-task scoped workload identity, short-lived credentials, operation-level authorization. No shared service principals, no standing production write access
Approval matrix — mandatory human approval for money, access, data egress, regulated decisions, and destructive actions
Agent registry and lifecycle control — every running agent, its owner, scope, supervisor, restart policy, and a disable path that survives redeploy and scheduler
Data contracts where data is in the path — because agents writing SQL against uncontracted tables is a control failure, not an analytics inconvenience
Acceptance evidence
A containment drill, run live, including a restart attempt — and it holds for a full cycle
Evidence that an unsafe action was correctly refused, not just that a safe one succeeded
Every write path produces an immutable audit record with trace correlation
Dry-run output for each destructive tool, reviewed before first live use
Environment provisioning meets its stated SLO from a cold start
Your operator runs the whole procedure once, unaided, and we watch
Six to ten weeks, working range $80k–$150k. Guardrails filter inputs and
outputs. No guardrail catches a plausible artifact sourced from the wrong place. This is the
engagement that makes agent output trustworthy at volume, and it is where the relationship
stops being replaceable by a product release.
Built
System context graph — a generated, continuously refreshed, machine-readable map of services, repos, pipelines, datasets, environments, dependencies, owners, SLOs, on-call, and cost centres. Built from IaC state, CI config, the cloud resource graph, catalogue metadata, git history, and incident records — never from a wiki somebody has to maintain
Execution memory — active plans with completion criteria, decision logs, progress files, quality grades, and structured handoffs, so a fresh session resumes instead of restarting
Provenance and postconditions — layer 0 of the evidence plane. Never accept a completion claim; assert independently that the end state holds, that inputs came from the authorised source at the expected version, and that prohibited side effects did not occur
Verification speed — test selection, parallelism, caching, flake quarantine, and merge-queue design, so time-to-verdict stays under the time an agent takes to produce the next change
Operator adoption track — the operators who own the workflow help author the golden suite, human judgment stays authoritative at named decision points, and override rates get acted on
One workflow in production — engineering-facing, internal blast radius, fast feedback
Measured, not asserted
Resumption success rate — the share of interrupted tasks a fresh session continues correctly without human re-briefing
False-completion catch rate — against traps we built into your golden suite
Time-to-verdict — baseline, target, and the flake budget that protects it
Cost per successfully completed task — the number a CFO can actually use
Override rate and week-four sustained usage — whether anyone still trusts it a month later
Candidate first workflows. Dependency and framework upgrades at scale;
test backfill and flake reduction; IaC drift remediation; data pipeline migration;
incident-context assembly; SQL or dbt model generation behind contract gates. All
engineering-facing, all with a buyer who feels the pain personally. Nothing touching
customers, money, or regulated decisions in engagement one.
— How we pick the wedge workflow with you
Ongoing
AgentOps
$15k–$40k per month, offered only once you have an agentic workflow in
production. We do not sell a retainer before there is something to operate — a subscription
with nothing running under it is a tax, not a service.
Model and provider change
Evaluation of every model or provider change against your golden suite before it
reaches your workflows. Models change under you whether or not anyone asked them to,
and a silent capability regression is indistinguishable from a bug you caused.
Evaluation and drift
Eval maintenance, new traps as new failure modes appear, long-run reliability review,
and data-quality gate operation on every agent-touched dataset.
Economics and incidents
Cost-per-successful-task management against a stated budget rule, agent incident review
with a named human owner, and the adoption and override review that tells you whether
anyone is actually using it.
Alternative entry
When data is the whole engagement
Data is usually a component of the rungs above. Occasionally it is the problem, and
starting anywhere else wastes your money. The standard door is still the runtime baseline —
this is the exception, not a second front door.
The signs
An agent programme is blocked specifically on grounding quality or confidently wrong answers
There is no test-data path, so nothing can be safely built — very common, very rarely diagnosed
An incident traced back to stale or mis-permissioned data
Agents are already writing SQL or pipelines against an uncontracted warehouse
The engagement
A data readiness baseline establishes what agents can actually see and whether retrieval
respects the permissions of the person asking. A foundation project then builds the
semantic and contract layer, quality gates, lineage, and the masked environment path
that unblocks everything else.
Working ranges: baseline $12k–$20k, foundation project
$30k–$55k, 90-day stabilisation $6k–$12k/month. Sold
alongside the ladder, not instead of it.
See the data foundation →
Catalogue
Modules, components, and integrations
The catalogue is what stops a repeated need becoming one-off custom work. Anything marked
destination is part of the published path and is built when a client pays for the
behaviour — never speculatively, and never implied as present capability.
Module
What it does
Rung
Status
Outcome ledger
Four-tier measurement spine wired to real telemetry; cost per successfully completed task
1
Available
Golden task harness
Client-specific task corpus with known-good outcomes and false-completion traps; runs in your CI
1
Available
Agent registry
Every running agent, owner, scope, supervisor, restart policy, disable path
The 70/20/10 rule governs every module: 70% supported core, 20%
configuration, at most 10% client-specific code. A permanent fork gets priced properly or
declined — quietly maintaining bespoke forks is how good delivery organisations become bad
ones.
Underneath all of it
The platform and data engineering we bring to the work
The agent-facing layer is new. What it stands on is not, and we do not treat it as an
afterthought — most of what makes an agent programme survive contact with production is
ordinary platform engineering done properly.
Not sure which one you need? That is the first meeting.
Bring the agent nobody owns, the pipeline nobody can reproduce, or the audit finding you
cannot answer. We will place you on the ladder and tell you what the next rung costs. If
the honest answer is that you are not ready for any of this yet, we will say that instead.
