Skip to content

The ideal Lytos team

“The repo is the project. Not Jira. Not Notion. Not the meetings.”

Agile didn’t fail. It solved a real coordination problem — and it worked for twenty years.

But it solved it by building a management layer above the code. Story points, sprint ceremonies, Jira tickets, burndown charts — a parallel world that tracks the project without touching it. A world where the people responsible for delivery spend a significant part of their time in tools that don’t know what a commit is.

The repo — where the project actually lives — became a detail. Something developers dealt with. Something managers projected onto a board elsewhere.

Lytos starts from the opposite premise.

Rules live in the repo. The issue board lives in the repo. Memory lives in the repo. The manifest — what the project is, what it isn’t, what quality means — lives in the repo.

Not because it’s clever. Because it’s where the work happens. An AI agent that reads .lytos/ before every session knows the project. A new team member who clones the repo has the full context. A decision recorded in an ADR is auditable two years later.

No external tool required. No sync needed. No drift between what’s written and what’s built.

This is the paradigm shift in one sentence: the speed of a builder, the quality of a lead dev, without the ceremony overhead.

Agile answered “waterfall plans too much — let’s ship fast.” Lytos answers “AI executes fast but without traceability — let’s govern without slowing down.” Not Agile 2.0. A new premise.

Most teams that adopted AI tools in 2024–2025 are in this situation: the AI is fast, the output is uneven, nobody really knows why a decision was made, and context disappears between sessions. If you recognize yourself in the middle column, that’s the problem Lytos solves.

Before AIMost teams todayWith Lytos
Project memoryIn heads, Confluence, meetingsThe AI knows nothing the next morning — same briefing every sessionIn the repo, reloaded automatically every session
AuditabilityTicket + PR — partialNobody can answer in code review: who generated this, with which model, on which prompt?Issue + ADR + commit — reconstructible two years later
Delivery speedLimited by human coordinationFast, but unpredictable — significant reworkFast and consistent — the frame prevents rework
Code qualityDepends on the human reviewerTwo developers, the same prompt, two different architecturesrules/ applied every session — consistency is structural
ConventionsVerbal, in Confluence, rarely enforcedReinterpreted at every session — or simply ignoredVersioned in rules/, enforced automatically
AI hallucinationN/AUndetectable until review — the AI invents decisions, ignores constraintsManifest + rules reduce drift — every session restarts from the same constraints
SovereigntyData locked in a SaaS vendorContext locked in the AI provider’s chat historyEverything in Git — you own it, portable, no vendor lock-in
Switching AI providersN/AChange Claude for GPT = all context lost, start overSwitch the engine, keep the frame — zero context loss
AI costN/AConstant rebriefing = wasted tokens every sessionStructured context = predictable cost, cacheable
ReproducibilityProcesses documented, hard to replayPrompts are throwaway — impossible to reproduce what was producedIssue + skill + rules = reproducible, transmissible workflow
OnboardingWeeks of context transferEach new AI session — or new team member — starts from zeroClone the repo, read .lytos/ — immediately operational

This premise has a natural consequence for how a team is structured.

The PO stays close to the user. They define what to build and why — outcomes, not tickets. They don’t write stories for developers who need hand-holding. They don’t attend standups. They know the user.

The Gouvernant owns the frame. They write the manifest, the rules, the architectural decisions. They define what “correct” means for this project, once, in the repo. Every AI session and every Builder operates inside that frame. The Gouvernant doesn’t need to be present for every task — the frame does the work.

The Builders direct execution. They write clear issues, give the AI the right context, review the output, and commit. Their value isn’t typing speed. It’s judgment — knowing when the AI got it right and when it didn’t.

That’s the team. Three roles. No Scrum Master. No ceremonies.

Agile roleStatusLytos equivalent
Product OwnerRefocusedPO — outcome strategist, close to the user
Lead DeveloperElevatedGouvernant — writes the frame, not the code
DeveloperRedefinedBuilder — directs AI, reviews, commits
Scrum Master→ FeatureProcess lives in the repo. Role is optional.
QA→ ToolingAbsorbed by rules/ + lyt lint
Agile CoachDisappearsThe method is in the repo, not in a consultant
Agile artifact / ceremonyStatusLytos equivalent
Jira / Linear tickets→ RepoIssues in .lytos/issue-board/
Confluence / Notion docs→ Repomanifest.md, memory/, ADRs
Verbal coding conventions→ Reporules/ — versioned, applied every session
Definition of Done→ Reporeview field in issue frontmatter
Sprint planningDisappearsPriorities set in the board, always current
Daily stand-upDisappearsBoard state in Git reflects reality
RetrospectiveDisappearsIssues and ADRs capture what matters
Onboarding sessionsDisappearsClone the repo, read .lytos/

A team of three to five people, working repo-first, can move faster and ship cleaner than a Scrum team twice its size — because there’s no management layer to maintain.

That’s not a feature of AI. It’s a feature of removing the gap between where the project lives and where the work happens.

The Builder is a high bar. Directing AI well requires domain knowledge, technical judgment, and the ability to evaluate output. Someone arriving at Lytos without that background will struggle. Two answers: either the Gouvernant’s rules are strong enough to compensate, or Lytos helps Builders develop judgment through tool feedback — a poorly structured issue is flagged with an explanation, not just blocked.

Governance must disappear into the gesture. If it feels like overhead, it gets bypassed. An issue drop that triggers the right gitflow, an issue that opens with the right context pre-loaded — that’s the central challenge of any repo-first workflow. The frame only works if working inside it is faster than working around it.