Agentic workflow supervision · Build & Run

An agent in production drifts. We watch yours

Data changes, models change, costs slip - and an agent that answered correctly in January answers wrongly in June without anything ever crashing. We supervise your agentic workflows continuously: cost per request, failure rate, drift in answer quality, compliance. A monthly subscription, not a project.

Discuss your project See the methodology
They trust us
Monthly subscription
€3,000

Excl. VAT, per month, starting price. The tier depends on how many workflows are supervised and at what volume.

Commitment
6 months, then monthly · 2 months notice
Setup
2 to 3 weeks of instrumentation before the first invoice
Talk about your run

We also supervise systems we did not build.

What we watch

  • Cost per request and per workflow, with threshold alerting
  • Failure rate, retry rate, end-to-end latency
  • Drift in answer quality, measured against a versioned test set
  • Guardrail activity: refusals, escapes, unexpected tool calls
  • Decision traceability and AI Act compliance log
  • Health of MCP connectors and RAG sources (freshness, coverage, errors)
  • A one-hour monthly review with your teams
  • Guardrails and test sets kept up to date as the system evolves

What you always have

  • A dashboard your teams can access
  • A written monthly report: costs, incidents, drift, decisions
  • SLOs defined with you, and measured
  • A named architect, reachable, who knows your system

What is not included

  • No new feature development
  • No managed services for your whole infrastructure
  • No level-1 end-user support
  • No 24/7 on-call by default - see the FAQ

The challenge

What breaks after go-live

A successful go-live is not an ending: it is the start of the period where things degrade slowly, without warning. None of these triggers a 500 error - which is exactly why they go unnoticed for months.

  1. 01 Cost per request triples after a model change and nobody sees it
  2. 02 Answer quality drops as the reference data ages
  3. 03 An agent gains broader permissions during an evolution, and keeps them
  4. 04 Guardrails are bypassed by phrasings nobody anticipated
  5. 05 The test set has not been updated since delivery and no longer detects anything
  6. 06 Nobody can reconstruct why an agent made a given decision three months ago

Our methodology

How we deliver

Architects with twenty years behind them, and a delivery process instrumented by AI: scoping, code review, quality and security tooled at every phase. Each step produces concrete deliverables, and we stay through the run.

01
2 to 3 weeks

Instrumentation

Before supervising, you have to measure. We instrument your workflows and define with you what counts as an anomaly - a threshold nobody set never fires.
Deliverables
  • End-to-end tracing of the agentic chains
  • Reference test set, versioned like code
  • Cost, latency and failure-rate thresholds set with your teams
  • Written and validated SLOs
  • Dashboard in place and accessible
02
Week 3

Going under supervision

Switch to continuous supervision. Alerting goes to your channels, not ours: you see what we see, at the same moment.
Deliverables
  • Alerting wired to your channels (Slack, Teams, PagerDuty or email)
  • Written escalation procedure with named contacts
  • First baseline measurement of cost and quality
  • Runbook of known incidents
03
Ongoing

The monthly rhythm

Every month: a written report and an hour of review. The report arrives before the meeting, so the hour is for deciding rather than discovering.
Deliverables
  • Monthly report: costs, incidents, measured drift, decisions to take
  • One-hour review with your teams
  • Guardrails and test set updated
  • Costed optimisation trade-offs (AI FinOps)
04
Every 3 months

The quarterly review

A step back onto the architecture itself: what should evolve, what costs more than it returns, what should be switched off.
Deliverables
  • Architecture and accumulated-debt review
  • Costed evolution recommendations
  • Compliance review and AI Act file update
  • SLO report for the quarter

Business value

What you concretely gain

Expected results

Drift shows before your users see it

Cost becomes steerable again

Compliance stays current

  1. 01

    Drift shows before your users see it

    A versioned test set, run continuously, catches a drop in answer quality while it is still fixable - not when the business reports complaints.

  2. 02

    Cost becomes steerable again

    Cost per request and per workflow, with threshold alerting. The gap between two implementations of the same function can reach a factor of twelve - but only if you measure it.

  3. 03

    Compliance stays current

    The traceability log and AI Act file are maintained as you go. A vendor questionnaire takes a day to answer, not three weeks of reconstruction.

  4. 04

    An architect, not a service desk

    The same person who designed or audited your system supervises it. No ticket climbing three levels before reaching someone who understands the architecture.

Frequently asked questions

Your questions, our answers

01 Can you supervise a system you did not build?
Yes, and it is common. We start with a diagnostic - two weeks to understand what exists and instrument it properly. Supervising a system you do not understand amounts to watching graphs without knowing what they mean.
02 Do you provide 24/7 on-call?
Not in the base subscription, and we would rather say so plainly than imply otherwise. Supervision is continuous, but human intervention happens in business hours with a committed response time. Extended on-call is contracted separately, based on your real SLOs - many organisations pay for it without needing it.
03 What happens concretely when an agent drifts?
The alert reaches your channels and ours at the same time. We qualify it, document it, and put a trade-off to you: immediate fix, temporary restriction of the agent scope, or rollback. The decision stays yours - we do not change a production system without your agreement.
04 Why a subscription rather than ad-hoc interventions?
Because drift is continuous and ad-hoc intervention always arrives after the fact. On-demand billing makes you pay for incidents; the subscription pays for their absence. If after six months you have had neither an incident nor a cost overrun, the subscription did its job.
05 What if we want to stop?
Six months initially, then monthly with two months notice. On exit you keep everything: instrumentation, test sets, dashboards, runbooks and documentation. Nothing is hosted with us in a way that would make you captive.
06 What exactly does the price depend on?
The number of supervised workflows, the request volume, and the SLO level agreed. €3,000 a month corresponds to a production system with one or two workflows and moderate volume. The tier is set after instrumentation, never before: costing supervision without having measured would be guessing.

Who is watching your agents right now?

If the answer is "nobody" or "we check the logs when someone complains", let us talk for thirty minutes. We will tell you whether supervision is justified at your stage.

Schedule a call See all our solutions