One request. A whole engineering crew.

Aim once. Land once.

Turn a clear request into a reviewable code branch.
Arrow plans, builds and proves the change before it lands.

Open Demo Explore the code

MODEL EXECUTION BY IBM Bob Shell

THE ARROW ARCH / your-next-feature WORKFLOW PREVIEW

01 → 05 FROM INTENT TO EVIDENCE

A complete orange trajectory from a request through an AI crew and verification checks to a target.

One direction. Every step accounted for.

VERIFY 04

“Done” is a claim. Here’s proof.

Arrow reads the diff and independently runs the checks. A worker’s summary is never enough.

Actual diff inspected
File scope validated
Required checks re-run

OUTPUTproof-report

Isolated workersIndependent checksA branch you can reviewWHY ARROW EXISTS

[ 01 / THE REASON FOR ARROW ]

Real problems.
A more deliberate system.

Built around the things developers actually say when AI work goes sideways.

RESEARCH THEMES/ paraphrased X Reddit Stack Overflow

It forgot the earlier decision.

The fix broke another part.

Two agents changed the same thing.

Context disappears

Earlier decisions get lost. The re-explaining starts again.

Workers collide

The same files. The same ports. Conflicting changes.

“Done” is not proof

A confident summary can still hide a failing test.

Project conventions drift

Wrong folders, wrong APIs, unnecessary dependencies.

Fixes create collateral damage

A repair breaks working code. The loop continues.

Humans still babysit the output

Reviewing everything becomes a second full-time job.

Real complaints → repeatable rules → safer executionAlso in the reports: oversized context, invented APIs, the wrong thing built, environment setup failures and skipped or repeated tool calls.

[ 02 / HOW IT WORKS ]

One prompt.
A whole crew behind it.

Each role exists to stop a different type of mistake. Every handoff has a purpose.

ONE REQUEST/ shared directionEXAMPLE REQUEST

“Add discount codes to checkout. Keep the existing payment flow intact.”

01 / PM

Understands the ask.

Turns your request into clear acceptance criteria.

Acceptance criteria

02 / ARCHITECT

Chooses the way.

Sets structure, file scope, dependencies and proof commands.

Bounded work packets

03 / WORKERS

Build small pieces.

One bounded packet. One isolated Git worktree per worker.

Scoped changes

04 / ORCHESTRATOR

Keeps control.

Owns state, order, processes, retries and the live ledger.

Controlled execution

05 / VERIFIER

Proves the result.

Reads the actual diff. Independently re-runs the checks.

Independent proof

Reviewable branch + proof trail READY FOR HUMAN REVIEW

[ 03 / YOUR REPO. YOUR RULES. ]

Onboard once.
Run with your rules.

Arrow learns your repository, your rules and the way your team works. That project context stays available for future tasks.

PROJECT MEMORY/ onboardingREPO → MEMORY → EVERY TASK

WHAT ARROW LEARNS

01Repository
02Company / team rules
03Folder structure
04Commands + versions
05Tech stack
06Coding conventions
07Approval rules
08Existing workflows

Future Arrow tasks New developer

PROJECT MEMORY

Keep the context. Skip the re-explaining.

Repo profile, commands, allowed structure, team conventions — including the rules you deliberately override.

A NEW DEVELOPER JOINS?

The same memory. A human head start.

The project context that guides agents can also explain where things live, how to run the repo and which conventions matter.

[ 04 / TRUST THE LANDING ]

Workers can say “done.”
Arrow still checks.

The model’s summary is never the final verdict. The control plane reads the actual diff and re-runs the proof.

Isolated work

Each worker edits its own Git worktree.

File scope

Changes stay inside the packet’s allowed files.

CHANGE REVIEWILLUSTRATIVE RUN

src / checkout / total.ts+3 −1

12export function total(items) {

13 − return sum(items);

13 + const subtotal = sum(items);

14 + return applyDiscount(subtotal);

15}

Changed files match allowed scopePASS

Required checks independently re-runPASS

Diff and proof attached to branchPASS

Proven. Ready for review.

Live ledger

Workers, processes, ports and ownership. Accounted for.

Independent proof

Arrow runs required commands before a packet can land.

Small fix or large feature.Same control system.Web appsMobileBackend / internalGame developmentAny codebase

Execution follows your repository’s toolchain and rules.

[ 05 / BEFORE YOU ASK ]

Questions are cheap.
Unproven code is expensive.

Still curious? Step through a sample request in the interactive walkthrough, or read the source.

Interactive walkthrough GitHub repository

No. Arrow handles bounded implementation work and proof. Humans still own product decisions, critical rule conflicts and review of what lands.

Your next change starts here