Maiestro Ensemble Agentic engineering for enterprise systems
Ensemble puts three agents on your enterprise software. The Analyst turns a request into a spec. The Builder writes the change. The Tester proves it on an isolated copy. Your people decide what ships.
01 Walkthrough
This is the path every change takes in Ensemble, shortened to under a minute. Press play, or jump to a chapter.
That was one request.
A request arrives in plain language. No template to fill in.
02 The agents
Each agent has its own workspace, its own rulebook and its own tools. Work passes between them as a sealed spec and an install package, so nobody builds on a guess.
Agent one
Turns a request into a spec a developer would sign.
Agent two
Writes the change, and only the change.
Agent three
Proves it on a copy of your system, then cleans up.
+2 The night shift
Every evening it re-reads the processes, code and data sources you connected and refreshes the search index. When nothing has changed it finishes in under two minutes and spends no model tokens.
At the end of the day it goes through each person's conversations, compares them with what is already stored, and keeps only what is worth remembering.
03 Deliverables
The workspace that writes the change also writes everything around it. The business unit gets a scope it can sign, testers get rules they can run, and engineering gets a quality report before anything ships.
Ask how anything in your system works: a process, a screen, a table, a rule. Every answer points to the code and data behind it.
You get: a clear answer with its sources.
Requirements, scenarios and open questions, written for the people who asked, in their language.
You get: a versioned analysis report and a scope the business unit can sign off.
Write test rules once. Ensemble runs the UAT scenarios live on the real screens of an isolated copy, as often as you like.
You get: pass or fail for every scenario, with evidence.
Static analysis of the code you already have: complexity, duplication, risky patterns and broken conventions, ranked by what matters.
You get: a quality report your team can act on.
Before anything changes: every screen, procedure, report and table it would touch, traced through the code.
You get: the full list, before anyone writes a line.
Everything the agents notice along the way, with location, severity and a suggested fix, tracked until it is closed.
You get: one list of what needs attention.
04 Knowledge
Ensemble does not work from a few files you paste in. It keeps a searchable map of every active process, the code behind it and the way your tables really connect.
Source code, stored procedures, screen and form definitions, data sources. Only from the systems you connect.
Personal data is masked before anything is stored: e-mail addresses, user codes, national ID numbers.
A model writes a short card for every unit: what it does, which tables it touches, which screens call it.
Cards go into a vector index. Search blends meaning, exact keywords and a reranker.
Many enterprise databases have no foreign keys. Ensemble mines the real relationships from the code itself.
“approval routed by amount”
Manager or finance director, by amount
Limit check before approval
Threshold read from the parameter table
Ranked by meaning, keywords and a reranker
1.2 M rows · statistics 2 days old
| Column | Type | Key / index |
|---|---|---|
| ID | NUMBER | primary key |
| CLAIM_NO | VARCHAR2 | |
| AMOUNT | NUMBER | indexed |
| STATUS | VARCHAR2 | indexed |
| CREATED_DAY | DATE | indexed |
Linked in code: EXPENSE_LINE.CLAIM_ID · APPROVAL_LOG.CLAIM_ID
05 Memory
Every user has a private memory. Preferences, decisions and past work come back on their own when they matter, on whichever engine is running.
Show amounts with their currency code.from your message · 3 Oct
Thresholds compare in base currency.expense-approval · 15 Oct
Rejected claims go back to the requester, not the team lead.expense-approval · 2 Sep
Spec v2 sealed. Package 128 passed 4 of 4.written by the system · 15 Oct
Use the old approval matrix.replaced on 12 Oct
Preferences and decisions are stored from the user's own words, with a link back to the message.
A decision is attached to a process or a data source and loads by itself when someone opens it again.
Memory never overrides the spec. A sealed spec always wins.
Every save shows a short notice with Undo. People can edit or delete anything on their memory page.
When something stops being true it is marked outdated, and both dates are kept.
06 Engines
Ensemble is not tied to one model vendor. The same agents, tools, rules and memory run on every engine, and you choose per person or per role.
Claude Code
Anthropic
Codex
OpenAI
OpenCode
Open source
Analyst · Builder · Tester
Tools · Rulebooks · Memory · Run log
Chosen per person or per role. Hosted or local.
Effort → model, per project: low · medium · high · max
Set a default engine for each agent and override it for anyone who prefers another.
Low, medium, high and max each point to the model you pick for that project.
When a new engine version arrives, Ensemble tries it against every model in use before switching. If one check fails, it stays on the version that works.
07 Safe release
Testing is one button. The rehearsal runs on an isolated copy with the same structure as production, and it cleans up after itself.
One button. Pick the scenarios you want to run.
The package is installed on a copy with production's structure and code.
A browser agent signs in as test users and works each scenario like a person would.
The package, the test data and every process the run opened are removed.
To the staging copy with one click. To production as a script a person reviews.
Agents never write to your production database. There is no tool for it.
08 Workflow
Every spec lives in a shared pool with its full history. People pick up each other's work without losing the thread.
Each seal creates a version. Changes are shown line by line, in place, the way a pull request shows code.
While a spec is sealed nobody else can take it over. Release it and a colleague picks it up with the whole conversation.
Work sits on a built-in kanban board and syncs with Jira and GitLab when you switch that on.
Attach a video of the problem. Agents can look at the exact moment and the part of the screen that matters.
Move between platforms and environments (development, test, staging and more) from the same session.
Version history · board item EXP-212
3
engines, one set of agents, tools and rules
40+
purpose-built tools across eight tool servers
0
writes to production by any agent
<2min
nightly re-index when nothing has changed
09 Questions
Something missing? Write to hello@maiestrotech.com and a person will answer.
No. It takes a request to a tested change. People answer the open questions, seal the spec and decide what goes live.
Enterprise systems built from processes, screens, stored procedures and source code. Database access today is Oracle, always read-only. Talk to us about other stacks.
On a Linux server inside your network. The workspace, the orchestrator, the index and the memory all stay there.
Only what the engine you choose sends to its model provider. Use a local model and nothing leaves.
No fine-tuning. Ensemble learns through its index and its memory, both of which live on your server and can be inspected.
Yes. Choose the engine per person or per role, and map effort levels to models per project.
The release button stays off. The Builder reads the results, fixes the change, and the Tester runs the scenarios again.
Bring one real request. We run it through Ensemble with you, from the first sentence to a tested package.