The engineering behind agent teams
Give specialized agents a job. Give every result an owner.
Senior engineering leadership directs agents that plan, build, test and document software. The review and release path is defined for the work, with an accountable human responsible for what goes into production.
How a team is organized
One senior engineer. A team of specialized agents. A gate before production.
The engineer owns the architecture and the result. Agents plan, build, check, test and document. An independent reviewer checks the work. For a narrowly defined class of production bugs in the published MDSi case study, additional automated gates run through the release and notification steps.
We run our own companies this way, before asking a client to.
- ArchitectPlans the change
- BuilderWrites the code
- ReviewerA different model checks it
- QATests behavior
- SecurityReviews risk
- DocsRecords what changed
In our own companies
Two leaders, one approach, applied to our own businesses first.
Tom rebuilt his firm, Elegant Software Solutions, around AI agent teams he directs. Ben founded More Gain Capital and now directs people, data and AI together across its companies. They work alongside each other: Ben sets the operating priorities and owns adoption; Tom owns the engineering architecture and the agent teams.
In progress: being implemented now
Your Peptide Brand: comprehensive AI automation
An active implementation within the More Gain Capital portfolio.
The science
Practices that make agent-built work safe to run.
Specialized agents, not one chatbot
Planning, building, reviewing, testing, security and documentation are separate roles with separate instructions, so each step is checked by something that did not write it.
Independent review
A reviewer checks the exact work the builder produced. The published MDSi bug-fix pipeline adds a three-model adversarial gate for its defined scope.
Tests before fixes
When something breaks, the pipeline writes a failing test first, then the fix, and the test stays in the suite so the same bug cannot quietly return.
An agreed record
The design identifies which agent actions, approvals, exceptions and outcomes must be recorded so the people responsible can inspect the work.
Reusable learning
Where it fits the system, fixes and decisions become searchable knowledge instead of living only in someone's memory.
Evaluation and drift checks
Decision-making workflows need expected results and checks for drift. We agree the measures before putting them into operation.
The art
The part no model supplies.
Judgment about what to build
The expensive mistakes are choosing the wrong problem or the wrong architecture. Senior people make those calls, with operators in the room.
Knowing how the business runs
Our leadership has run routes, call centers and field teams. We design for the technician in the driveway and the manager at month-end.
Taste and restraint
Not everything needs an agent. Sometimes the right answer is a report, a form or a better process, and we will say so.
Agent teams we build for you
Agents that work inside your business, on your permissions.
These examples draw on work delivered by Elegant Software Solutions and on internal operating systems. The scope, permissions, records and approval points for your business would be designed with you.
Executive chief-of-staff agents
Triage mail, prepare briefings and track commitments across your accounts, with approval rules for anything sent or spent.
Customer-service coaching agents
Turn calls and texts into coaching for reps and technicians, with payment-card data redacted before analysis.
Sales and quoting agents
Turn a forwarded email or a voice note into a structured quote, logged for review.
Finance and back-office agents
Reconcile the books, prepare invoices and report on the results in your accounting system.
Engineering agents
Research production errors, propose fixes with tests and move them through review.
Knowledge and retrieval agents
Find answers in your own documents and articles by meaning, using semantic search.
Where the models run, how your data is handled and what we are not certified for yet are all on our security and trust page.
Stop piloting. Start shipping.
Bring us the workflow that eats the most hours. We'll scope an agent team for it and tell you honestly if an agent isn't the answer.