The Core Thesis: Context is Control

ChooChoo's mission is not to build better agents — it is to build the railway they run on.

AI agents are increasingly capable, but without structure they are "loose cannons": they hallucinate internal APIs, bypass team conventions, and act without the institutional knowledge your human engineers take for granted. The answer is not to intercept every token — it is to govern at three points: control the Input Context (what the agent knows at startup), observe the Output Trace (what the agent actually did), and continuously Evaluate (how well it followed the rules).

This is context engineering as a governance mechanism.

---

AI coding agents are powerful but fragile. Three compounding problems make most agent deployments brittle, unpredictable, and impossible to improve systematically. For a high-level view of how ChooChoo addresses these challenges, see the Architecture Overview.

1. Context Amnesia

Every agent session starts blank. Your agent doesn't know which files are off-limits, what decisions were made last week, why your auth system is structured the way it is, or what the team agreed about the schema migration. This is Context Amnesia: the loss of institutional memory between sessions.

ChooChoo's answer: Context Compilation. choochoo context generate scans your codebase and produces a structured AGENTS.md delivered to every agent at session start. Combined with Agent Traces, agents can retrieve past decisions and avoid repeating mistakes.

2. Ungoverned Agents

AI agents are autonomous by design — they read files, modify code, execute commands, and hand off to other agents without asking for permission. That autonomy is the point. But without a governance layer, you cannot answer: What did my agents do this week? Did they stay within their declared scope? Were any changes high-risk? Who approved them?

You wouldn't deploy a human engineer without code review, access controls, and an audit trail. AI agents are no different — they just move faster and make more changes per hour.

When multiple agents work in the same codebase, the problem compounds. Without declared boundaries, agents have no shared understanding of what they're allowed to touch. Without traces, there's no accountability for what they did. Without risk scoring, every change lands with equal weight — a typo fix and a schema migration treated identically.

ChooChoo's answer: Agent Rails + Schema Validation + Audit. Agents declare explicit rails and boundaries — what they can read, modify, and execute. choochoo validate enforces ODCS, ODPS, OpenAPI, Arazzo, GraphQL, AsyncAPI, and AI System Card schemas before changes land. Every action is traced and risk-scored; high-risk changes require human approval before they proceed. Governance failures surface during development — not in production, and not in an audit.

Good governance also unlocks second-order benefits that matter across the whole organisation:

• Cost governance — See exactly what your agent fleet is doing and what it costs. Track spend per agent, per team, per change type. Approve expensive operations before they run. Identify runaway usage before the bill arrives.
• Better intel — Engineering leadership gets real visibility into AI adoption: which tools are in use, what they're changing, where quality is trending. Decisions backed by trace data, not anecdote.
• Cross-team coordination — Security sets policy once in choochoo.toml. Legal's compliance requirements are baked in at the schema level. Engineering teams inherit governance automatically — no per-PR negotiation, no siloed interpretations of the rules.

3. No Feedback Loop

You can't improve what you can't measure. Most teams have no idea whether their agents are getting better or worse week-over-week. There's no harness to benchmark agent quality, no signal to fine-tune prompts, no data to train specialized models on.

ChooChoo's answer: Trace → Benchmark → Optimize. Every trace ChooChoo records is a data point. Traces feed evaluation harnesses (SWE-Bench, ITBench, custom task suites), surface quality scores in The Station dashboard, and provide the signal needed for prompt optimization and eventual fine-tuning. See the Roadmap for timeline.

What This Enables

Companies are hesitant to let agents loose on their codebases — for good reason. ChooChoo provides the safety harness that makes aggressive agent adoption safe:

• Governance-as-Code via choochoo.toml — policies are versioned, reviewed, and enforced like any other source code.
• Context-as-a-Service via AGENTS.md — every agent gets a structured world model for your specific repository, not generic training data.
• Continuous Verification via trace-backed evaluation — governance is proven by benchmark data, not assumed.
• Fleet Intelligence via The Station — cost, quality, and compliance dashboards across your entire agent fleet, visible to every team that cares.

ChooChoo follows a Bring Your Own Agent (BYOA) model. It standardizes the context format so the same policies, boundaries, and compiled context work across Claude Code, Cursor, Gemini CLI, OpenCode, and Codex. You pick the agents; ChooChoo ensures they all operate under the same governance rules.

Get Started

Ready to bring context engineering to your agent harness? Follow the quickstart guide to validate your first project in minutes.

• Installation — Install the ChooChoo CLI globally.
• Quickstart — Initialize a project, validate schemas, and connect your coding agent.
• Architecture — How The Engine, The Map, and The Station implement this vision.