Agents don’t need better models. They need deterministic tooling to execute across the code estate.
Today the default stance is a single agent working against a single repo.
People say “just let the agent do everything.” In practice that’s one agent in one repo editing code directly, and every change runs through inference. That works fine for one repo. The question is how it holds up when you’re running many agent sessions across many repos.
Default: one agent session per repository
Five repos, five sessions, five different diffs. Full inference every time — and no two land the same.
Every business problem has a shape, and some shapes are harder for coding agents than others.
Changing a type from X to Y is a small puzzle piece, and an agent can fit that one on its own. A Spring Boot migration is not a puzzle piece. It looks like the drawing below, and we keep handing it to agents whose only tool is inference, then telling them to go figure it out one token at a time.
Match the tool to the shape of the problem.
Deterministic tools are good at structural changes. The model is good at open-ended reasoning. Small changes can get by on the model alone, but the more complex the task, the more it needs deterministic tools underneath it. That’s what it takes to scale.
Serious people don’t doubt capability.
Serious people don’t claim consistency.
Deterministic change and agentic change look like they’re competing. They’re not. One gives you consistency and the other gives you capability, and at scale you need both. Instead of one agent per repo, picture one agent finding the right recipe, and the recipe making the change across every repo.
The inner loop is the recipe and its unit tests. The outer loop is an ever larger group of repositories. You pay a fixed cost up front, then the same change lands across the whole fleet. The inner loop costs the same whether you have ten repos or ten thousand. That’s herd safety.
Same shape, two modes. The agent’s skill is choosing which — the recipe when the change has a shape, the model when it doesn’t.
Moderne: one recipe, every repository
The same diff everywhere. The inner loop is the recipe and its unit tests; the outer loop is the repositories. Cost is constant irrespective of repo count — herd safety.
When a recipe breaks, it fixes itself once and runs everywhere.
A few agents can share one changeset across every repo. A bug in the recipe lives in one place, so the agents fix it once, re-run it, and the whole fleet heals at the same time.
Agents join one prepared changeset
Many agents, one changeset to review. Fix the recipe once; the whole fleet re-runs clean.
Agent tools, answered.
For one repo, yes. Across your whole engineering org, no, not without burning enormous token budgets on tasks an LLM was never designed for: deterministic, fleet-wide structural change. Moderne is the harness underneath the agent so the work scales, and your agent stays where it is.
It shifts work from GPU inference to CPU precomputation. Trigrep eliminates follow-up reads with symbol-aware search, Prethink generates context before the session begins, and for migrations, one recipe replaces hundreds of thousands of tokens of agent guesswork. A Java 8 to 25 migration drops from 61 million tokens to 30,000.
Moderne doesn’t replace your agent, it works alongside Claude Code, Cursor, Copilot, Codex, Windsurf, Cline, and any MCP-compatible agent. The agent you use today gets faster, cheaper, and more reliable without switching tools.
Through the Model Context Protocol, available in our CLI and SaaS product. You connect it like any other MCP server.
Trigrep is a symbol-aware, trigram-indexed code search built for agent workflows at enterprise scale. Its index is built from the LST, so results include type and symbol data. When an agent searches for a class or symbol, Trigrep returns semantically precise results in under a second with no follow-up reads required.
Prethink is a precomputed context layer that runs before an agent session. CPU-only static analyzers generate structured knowledge about architecture patterns, dependency graphs, and service boundaries, then make it available to the agent as context.
Recipes are deterministic, LST-powered transformations that agents call instead of writing ad-hoc code. Rather than grepping, reading, and editing thousands of files, an agent reaches for one of more than 7,200 recipes that execute correctly and consistently at scale.
Through governed, synchronized, deterministic execution across thousands of repositories at once. Organizations like Boats Group have executed changes across 300+ repositories in a single operation with full audit trails and no partial states.
Run the benchmark. See the 61-million-token difference.