The control layer for agentic coding

Lay the rails once.
Every agent runs them.

Same runbook for every engineer, every AI agent, every time. Scroll the rail — every phase, what the agent does, where a human must clear it.

01 / 07
Ingestion fixed

Jira, Sentry, Figma, a Slack thread — normalized into one run state. The only phase that can't be turned off.

scroll ↓

23:47 · nobody was watching

The agent shipped the wrong code at 11 pm — and nobody saw it until the customer did.

No checkpoint.No record.No way back.

That is the gate RailRun makes unskippable.

every run is a data point

Three engineers. Two agents. One track.

Claude Code and Cursor — different hands, identical rails. And every run on that rail is a data-collection point — drift, artifacts and activity stream straight back into the shared brain.

CC Jamie · Claude Code
Cu Sara · Cursor
CC Lee · Claude Code
Shared brain collecting · drift · artifacts · activity
how the software actually works

One loop. It compounds every cycle.

Loop primed.

Scroll — watch a run task move through the system.

where the magic happens

The agent hits a gate. A human clears it.

The one thing autonomous agents structurally can't show you — a stop.

Run task · TEAM-417 · live runbook v4 · locked at run start

Execution log

14:23:01 [agent] decomposing ticket via Claude…
14:23:04 [agent] scope manifest · 3 modules touched
14:23:05 [rail] phase SCOPE complete
14:23:08 [agent] drafting plan · MVP / MVP+ / MVP++
14:23:11 [rail] CHECKPOINT — agent halted
14:23:11 [agent] blocked · cannot self-approve

Tech lead · approval required

● checkpoint pending

Plan → Build

Agent proposes MVP+: 3 files, ~120 LoC, new migration. Authored by AI · run #173 · runbook v4.

Approve & release
Send back
gate: CLOSED · agent cannot proceed until a human clears it
▸ replay
runbook · design once

You draw the rail. On a canvas. It versions itself.

Drop a node — the lane reflows, the version ticks, the audit log writes. Nothing silent.

android-bugfix · project runbook self-improves 100% v4 · locked at run start
node library
unit testtool
scope promptprompt
compliance scanSOC2
human gategate
Ingest run-loadertool
Scope impact mapprompt checkpointgate
Plan MVP ladderprompt checkpointgate
Design figma→codetool
Build agent difftool
Test unittool securitytool
Ship PR opentool
audit · node added → Test / compliance scan (SOC2) · v4→v5 · by tech-lead · self-improvement applied · logged
compliance scanSOC2
▸ replay
run task · the work unit

Every run task, every agent, one board.

Each card is a run on the rail. AI-vs-human stamped on every move.

Project board · live drag = state transition · audited
Scope 1
AITEAM-417 · crash on cold start
Plan 1
humanTEAM-412 · gate review
Build 1
AITEAM-409 · settings migration
Ship 1
humanTEAM-402 · PR #168 merged
▸ replay
the collaboration layer

A shared brain. Real-time collaboration.

Every engineer's agent collaborates through one shared knowledge base — reading context in, writing learnings back, live. Seniors review what the agents and operators actually did.

Shared brain live collaboration
Team brain shared memory · evidence · learnings Jamie · Claude Code Sara · Cursor Lee · Cursor Ana · Claude Code

real-time collaboration · reads context in · writes learnings back

Senior review · operator performance
Jamie · Claude Code 88 approved 14
Sara · Cursor 71 sent back 3
Lee · Claude Code 94 approved 9

checkpoint decisions · diff authorship · AI-vs-human · audited

the engine room — evidence, not intuition

Evidence flows back. The rail gets sharper.

A ritual miner watches every run. When an off-script step repeats, it proposes an evidence-backed runbook patch — a human approves, it forks in, the version bumps. In-flight runs are untouched.

knowledge base
27 folders architecturestandards incidentsdecisions
→
ritual miner · drift
off-script repeat EMPTY_PHASESTACK_MISMATCH KB_GAP
→
proposal · P4
evidence-backed patch v4 → v5diffed scoped
→
human · P5
approve → fork ↺ audit loggedin-flight safe
P4 mine proposals — surfaced only once an off-script step repeats across runs; nothing applied automatically P5 apply on approval — forks into this project's runbooks + bumps the version; in-flight runs unaffected
ready when your team is

Lay the rails for your team.

One runbook, every AI agent, every engineer. Paste one key into your agent and run your first task in minutes.

For every agentic-coding startup — the workflow, audit, and team layer you didn't have time to build.

Start free → Sign in

Free forever up to 3 engineers · $9/seat from seat 4 · no credit card

CC Claude Code Cu Cursor

RailRun
RailRun