The software team
Build a software team with planner, coder and reviewer AI agents sharing a repository, scoped shell access and a virtual desktop for testing.
Scenario
You want software built — your product, an internal tool, the automation that glues your business together. One strong agent is a pair programmer; a team of agents with distinct roles, a shared repository, and a review gate is closer to an engineering org. Neuralis gives the team the same things a human team needs: the codebase as shared memory, separate seats with separate permissions, a machine to test on, and a schedule.
The team
| Agent | Job |
|---|---|
| Architect | Plans features, splits work, keeps the design notes current |
| Coder | Implements against the plan — reads, writes, runs the code through the policy-gated shell |
| Reviewer | Reads diffs hostile-first: correctness, security, the project's own rules |
Fan-out work — "explore this library", "find every caller" — goes to delegate sub-agents, whose read-only transcripts stay attached to the parent conversation.
The wiring
- The repository is a mounted, synced source. Code becomes semantically
searchable memory ("where do we validate uploads?"), and — the quiet
superpower — the repo's own
.claude/folder, its rules and skills, loads as a source package: the project's conventions ride with the project, the same files your human editors use. - The shell is the
executetool — sandboxed, environment-sanitized, path-policy-gated per source, so the coder builds and tests inside exactly the directories it is allowed to touch. - The machine desktop is the QA bench: run the app, click through it in a real browser, capture screenshots and recordings into the conversation.
- You watch from the terminal — your own PTY
rooted in the same source, because trust is good and
git diffis better.
The rhythm
- On demand — the feature loop: you brief the Architect; the plan lands in
data://design/; the Coder implements; the Reviewer gets the diff in a fresh conversation and files findings. - Nightly — Reviewer: sweep the day's changes against the project rules and leave a morning report.
- Weekly — Architect: dependency and drift review; design notes updated to match what the code now actually does.
Governance
Code ships consequences, so the seats are unequal on purpose: the Reviewer
holds read-only file access — it cannot "fix" what it judges; shell access
is per-source and path-gated, with host-level paths reserved to the admin
tier and above; anything destructive pauses on a Balanced profile for approval;
and conversations, runs, and shell calls are all in the audit log. Merges
and releases are yours.
Extend it
Add seats the way a team grows: a docs agent keeping READMEs honest, a triage agent working the issue queue, a release agent drafting changelogs. And the loop closes pleasingly: the software team can build your own packages — extending the very platform it works in.
The research desk
Assemble an AI research desk with MCP data sources, scheduled briefings and searchable notes, while people retain judgment and consequential decisions.
Connect anything
The capability pool: MCP servers, imported skills, public APIs, and your own packages — four on-ramps to one contract.