AGMENAG-men
Custom Agentic Systems
Agmen is Latin for a column on the march, from the same root as “agent.” Our systems move specialized agents in formation, under senior direction.
Production agentic software, built around your operations.
Is this for you?
- Operations-heavy businesses: intake, quoting, claims, billing, scheduling, reconciliation, reporting
- Workflows that cross several systems, such as a CRM, an accounting system, and a practice-management tool
- Teams without in-house engineers who need a production system, not a prototype
- Companies whose rules live in people’s heads and spreadsheets
The situation
The work runs across people, spreadsheets, and systems.
Your operations depend on work that moves between several systems and several people. Someone reads, checks, copies, and reconciles. The rules they follow are rarely written down in one place.
A prototype can show that an agent handles the familiar cases. A production system has to handle every case, write to the right record, and leave a trail your team can audit.
The question is no longer whether AI can do parts of the work. It is how to build a system your business can run, measure, and trust with its records.
Our approach
Senior engineers direct the agents. People approve the decisions.
We start from an agreed architecture, usually a System Blueprint: the boundaries, the rules, and one authoritative source of truth for each record. Senior engineers then direct specialized agents through bounded pieces of the build, and review each piece before it is accepted.
In the finished system, every output passes validation gates before a person sees it. Each agent works within explicit permissions, and consequential actions wait for human approval.
The system records the input, decision, correction, outcome, and cost of every run. That record becomes your own evaluation data, and it is how each later improvement is measured.
Use cases by industry
Where this service fits.
Typical applications across industries. They show where the service applies, not past client work or results.
- Accounting firm
Monthly close prepared across client books
Agents pull bank and card activity, match it to entries in the accounting system, and draft adjusting entries with the supporting documents attached. A staff accountant reviews each client’s reconciliation, and a manager approves every entry before it posts. Unmatched items go to the assigned accountant with the reason and the candidate matches shown.
- Logistics
Shipment exceptions handled across carriers and billing
The system reads tracking events from carrier portals and matches each delay or damage report to its shipment record. Agents draft the customer notice and any billing adjustment, with the transport management system as the source of truth. A dispatcher approves every customer message, and an account manager approves each credit before it reaches billing.
- Manufacturer
Quote requests turned into checked, priced quotes
Agents read each request for quote, match parts against drawings and the ERP item master, and draft a quote using the current costing rules. Engineering reviews any nonstandard part, and a sales manager approves every price before the quote reaches the customer. Each quote, correction, and won or lost outcome is recorded for later measurement.
- Clinic group
Referral intake and scheduling across locations
Referrals arrive by fax, email, and portal with uneven detail. Agents read each referral, check required fields and coverage, and prepare the patient record and a proposed appointment in the practice management system. Intake staff approve every record and booking, and sensitive records stay within the permissions agreed in the Blueprint.
- Construction
Pay applications checked against contracts and progress
Subcontractor pay applications are compared with the contract schedule of values, approved change orders, and site progress reports. Agents prepare a reconciliation that explains each difference and attach a draft approval for the project manager. The project manager and the controller both approve every payment before it reaches the accounting system.
- Wealth management
Client onboarding packets assembled and checked
Agents gather each new client’s onboarding documents, check them against the firm’s account-opening checklist, and prepare account forms from the CRM record. Missing or conflicting details go to the client service associate with the gap named. An advisor approves every form before it is sent, and no account record changes without that approval.
- Hospitality
Group booking requests quoted across systems
Group requests arrive by email and web form with dates, room counts, and event needs. Agents check availability in the property management system, apply the rate rules, and draft a proposal in the sales CRM. A sales manager approves every proposal and rate before it reaches the guest, and the booking is written only after acceptance.
- Energy services
Field inspection reports turned into records and invoices
Technicians submit inspection reports with photos and meter readings from each customer site. Agents extract the readings, update the asset record, flag readings outside the agreed limits, and draft the invoice from the service agreement. A service manager reviews every flagged reading, and the billing lead approves each invoice before it is issued.
What you receive
The application and its agents
Production software with agents that do bounded work: reading documents, preparing decisions, drafting, reconciling, and updating your systems.
Integrations with your systems
Connections to the tools you already run, with one authoritative source of truth for each kind of record.
Validation gates
Automated tests, evaluations, and checks on every output before a person reviews it.
Approval gates and permissions
Explicit limits on what each agent may do, and human approval on every consequential action.
Monitoring
Cost, quality, and outcome tracked for every run, with exceptions routed to the right person.
Data you keep
Every run records the input, decision, correction, outcome, and cost, which becomes your own evaluation data.
Documentation and handover
Runbooks and architecture notes your team or ours can operate from.
How it works
Architecture
Senior engineers define the system, its boundaries, acceptance criteria, and source of truth, usually in a System Blueprint.
Execution
Specialized agents do bounded work in parallel: analysis, implementation, testing, migration, documentation.
Validation
Automated tests, evaluations, and checks score every output.
Approval
People approve every consequential decision. Authority is never delegated to an agent.
Production
Accepted work ships, and every correction improves the next run.
How success is measured
The measures your approver signs.
Each measure goes into the acceptance criteria with its test data, threshold, and the person who checks it.
- Outcomes accepted against the Charter
- Each outcome in the Acceptance Charter is tested as written, on the agreed data, by the agreed person. Results are reported criterion by criterion.
- Errors caught by validation gates
- Whether the automated checks catch the errors a reviewer would catch. Known historical errors are run through the gates, and any miss is reported.
- Approval coverage
- Whether every consequential action waited for its named approver. Run records are audited to confirm that no payment, commitment, or record change bypassed approval.
- Consistency with the source of truth
- Whether each record agrees with its authoritative system after a run. Reconciliation checks compare the records across every integrated system.
- Cost per accepted outcome
- What each accepted outcome costs, including model use and infrastructure. It is calculated from run records, which also show where later work can lower it.
Where care is needed
What we watch, and how it is handled.
- Rules held in people’s judgment
- Many rules live in the judgment of experienced staff. We write them down with those people, and the owner approves each rule before agents apply it.
- Permissions on live systems
- Agents work inside explicit permissions on every system they touch. A principal engineer holds authority over security and every irreversible change.
- Records held in two places
- When two systems hold the same record, errors can travel between them. We name one authoritative source for each record before integration begins.
- Model change over time
- Models are updated and replaced during a system’s life. Run records give a measured baseline, so any model change is tested against the agreed criteria before release.
- Operation after handover
- The system should be operable by whoever runs it next. We hand over runbooks and architecture notes, and walk that team through daily operation and exceptions.
Who does what
Your team decides. We engineer.
Your team
- Name an owner who can sign the Acceptance Charter and make scope decisions
- Confirm which system holds the authoritative version of each record
- Provide access to the systems in scope, under agreed permissions
- Explain the business rules, including the exceptions staff handle by judgment
- Approve consequential actions and accept delivered outcomes against the Charter
Sophrono
- Hold authority over architecture, security, and every irreversible change
- Direct the agent team and review its work before acceptance
- Build the integrations, validation gates, permissions, and monitoring
- Run the agreed acceptance tests and report each result
- Document the system and hand it over to your team or ours
At the end
The decisions you make next.
The service ends with evidence and a choice. Each option is yours, and none is assumed.
Run it with your team
Your team operates the system from the runbooks and architecture notes. The run records keep measuring every outcome.
Improve one measure at a time
Iteration Sprints make one measured change per sprint, and your approver accepts or rolls back each release.
Have us look after it
AI Stewardship keeps the system monitored and current, with people approving each release. Its tiers are Foundation, Continuity, Operations, and Custom.
Extend to the next workflow
The architecture, integrations, and run records give the next workflow a starting point. New scope goes through a change order or a new Acceptance Charter.
Before we start
What to have ready.
- A System Blueprint, or the workflow map and rules one would produce
- A decision on who approves payments, customer commitments, and record changes
- Representative examples of the work, including exceptions and edge cases
- A list of the systems involved and who administers each one
- A view on where the system should run: a hosted service, your cloud, or your hardware
Related services
The agent team
Senior engineers direct. Specialized agents execute.
Agents work in parallel only where it is safe, and as many as the work can divide cleanly. Every output passes automated validation before a person reviews it.
- Principal engineerHolds authority over architecture, security, source-of-truth changes, acceptance, and anything irreversible.
- Requirements agentTurns business goals into structured requirements.
- Implementation agentsBuild independent pieces of the system in parallel.
- Test-generation agentWrites tests from the specification and its edge cases.
- Validation agentChecks every piece of work against the acceptance criteria.
- Security agentChecks secrets, permissions, dependencies, and unsafe actions.
- Integration agentChecks that independently built pieces work together.
- Documentation agentKeeps technical and operating documentation current.
What “accepted” means
Measured against criteria you agree to in advance.
- Every outcome is listed in the Acceptance Charter you sign before the build: its test, its threshold, the data it is tested on, and who runs the test.
- The acceptance process and review period are finalized in a Master Services Agreement.
- New or changed criteria go through a change order, so scope and price stay clear.
Request a build proposal
Describe the system you want built.
If a Blueprint or pilot comes first, we say so. Otherwise we reply with what a proposal needs.
- A senior engineer reads every request
- A reply by email with the next step
- No obligation until scope and price are agreed
Not ready to scope this? Ask an engineer first: a free 15-minute call that names the agentic systems that could fit.
Metron ariston · Measure is best.
Reliability is engineered.
Context, tools, permissions, and validation determine production reliability. Custom Agentic Systems is one of the ways we hold to it.
Read Canon VFree template
See exactly what “accepted” means.
The sample Acceptance Charter shows how every outcome gets a test, a threshold, test data, and an owner before a build begins.
- Criteria written as tests
- A clear acceptance window
- Change orders for new criteria
Sample Acceptance Charter
Preview it firstWhere this fits
- Learn Research
- Explore free AI Workload Evaluation
- Try one workflow Diagnostic
- Plan and validate System Blueprint
- Build Custom Agentic Systems
- Improve and operate AI Stewardship
Not ready yet? System Blueprint. After this: AI Stewardship or Agent Assurance.
Questions
Which AI models do you use?
Whichever clears your quality bar at the lowest cost, task by task. That can be leading coding agents such as OpenAI Codex and Claude Code, or self-hosted open-weight models where privacy, control, or cost call for it. Each choice is checked against the agreed criteria on your work.
Who owns what you build?
Ownership and licensing terms are finalized in a Master Services Agreement.
Do the agents act without approval?
Only within limits you approve in advance. Consequential actions, such as payments, commitments to customers, or changes to records of truth, wait for a person.
How long does a build take?
It depends on scope. The System Blueprint gives a schedule alongside the fixed price, before you commit to the build.
Can it run in our own cloud or on our own hardware?
Yes. Systems can run in your cloud account, on premises, or both. The Self-Sovereign AI section covers the options.
What happens after launch?
You can run it yourselves with the documentation provided, add Iteration Sprints (Iterum) for improvements, or have us look after it through AI Stewardship (Custos).
How do we choose between fixed price and hourly?
A fixed price suits a scope defined in a signed Acceptance Charter. Hourly work, at $375 an hour, suits scope that is still evolving.
Do we need a System Blueprint first?
Usually. The Blueprint defines the architecture and acceptance criteria that the fixed price is set against. If your scope is already well defined, we agree what the build needs before it starts.
How many agents work on a build?
As many as the work can safely use: fewer on tightly coupled work, more on work that separates cleanly. A principal engineer reviews the work, and success is measured by accepted output.
Where do the run records live?
With the system, in the environment we agree with you. They become your evaluation data for measuring later changes.
Can we start with one workflow before the full build?
Yes. A Production Pilot runs one workflow in production conditions against criteria you sign, and its evidence informs the full build.