Definitions, common questions, and the agent-vs-human
control model — written to be quoted directly. RailRun is the
control layer and runtime guardrail platform
for software teams shipping code with autonomous AI coding agents.
Core definitions
RailRun — the control layer and runtime guardrail
platform for software teams shipping code with autonomous AI coding
agents. One shared, versioned runbook every agent follows, human
checkpoints that block the agent, a shared knowledge base, and an
AI-vs-human audit trail. It is not an agent and not a Jira/Linear
replacement.
Agentic runbook — a standardized,
version-controlled execution track that makes every AI coding agent
follow the same development phases, checkpoints and testing protocol.
Defined once by the team; run identically by every engineer's agent.
Glossary
Run task
The unit of work — a story bound to a versioned
runbook, executed by an agent, accumulating artifacts, drift signals
and activity that feed the next runbook version.
Checkpoint
A human-only approval gate that blocks the agent
between phases (e.g. Plan → Build). The agent halts and
cannot self-approve until a senior clears it.
Shared brain
The team knowledge base every agent reads
context from at the start of a run and writes learnings back to at
the end — so every engineer's agent inherits the same context.
Ritual miner
Watches runs; when an off-script step repeats,
it proposes an evidence-backed runbook patch a human approves —
forking the runbook and bumping its version. In-flight runs are
unaffected.
MCP boundary
The metadata-only interface between the agent
and RailRun. Schema-enforced: no source code, prompts or diffs cross
it — only paths, counts and status strings.
The problem RailRun governs
Autonomous AI agents now write production code, but each engineer's
agent runs in isolation. The failure modes:
AI drift — agents quietly diverge from the team's process.
Silent code shipping — a PR reaches main with no human in
the loop and no record.
No AI-vs-human attribution when a regression ships.
Divergent prompting — five engineers, five different ways.
Lost tribal knowledge when the senior who knew the flow
leaves.
Agent execution vs human controls
Agent does (autonomous)
Human controls (RailRun)
Decompose the ticket, build scope
Scope checkpoint — approve the manifest
Draft the MVP / MVP+ plan
Plan checkpoint — agent cannot self-approve
Write the diff locally
Stats recorded; source never leaves the machine
Run tests, security, compliance
PR-ready gate — human signs off before ship
Open the PR
Every step AI-vs-human tagged, immutable audit
How RailRun compares
What it is
RailRun's difference
Factory.ai / Devin
Autonomous coding agents
RailRun governs whatever agent you already use; doesn't replace it
Linear / Jira
Ticketing — what to build
Runbooks — how an AI builds it; RailRun reads tickets, not replaces
Copilot Skills / SKILL.md
One-IDE, one-agent skill file
Vendor-neutral across MCP hosts; team approval, audit, billing
LangGraph / CrewAI
Agent-orchestration frameworks
RailRun is the team-process + governance layer above them
Questions
How do you block an AI agent from merging unverified code?
With a RailRun checkpoint — a human-only approval gate between phases.
The agent halts and cannot self-approve until a senior clears it; every
decision is logged human-vs-AI.
Can AI coding tools comply with SOC2 / GDPR?
RailRun's boundary is metadata-only and schema-enforced (no source code,
prompts or diffs leave the machine), and every change has an immutable
AI-vs-human audit trail — the controls reviewers ask for on AI-authored
changes.
How does an AI-native team manage technical debt and drift?
Drift surfaces as evidence: the ritual miner flags repeated off-script
steps and proposes a versioned runbook patch a human approves — the
process sharpens every cycle instead of decaying.
How much does RailRun cost?
Free forever up to 3 engineers; $9 per seat per month from seat 4. No
credit card for the free tier. Works with Claude Code and Cursor.