Adaptive parallel execution for Codex

Run more.
Break less.

AtomLane proves which commands can run together, then sizes concurrency for the machine actually doing the work. Long runs stay visible. Time savings stay measurable.

macOS Stable Windows Preview v0.17.0 MPL-2.0

Install AtomLane View source ↗

AtomLane live progress showing active tasks and estimated time saved — ACTUAL UI CAPTURE

One decision loop

Parallelism is compiled,
not guessed.

AtomLane sits between a coding agent’s plan and local execution. It accelerates only the work whose dependencies, effects, outputs, and resource needs can be bounded.

01 / ASSESS

Find real work

Recognize tests, builds, containers, batch jobs, and long-running scripts worth evaluating.

02 / PROVE

Compile atomic lanes

Model dependencies and reject overlapping writes, unsafe state, and unclear execution boundaries.

03 / SIZE

Match the host

Allocate CPU, memory, workers, and accelerators across every active AtomLane session.

04 / RUN

Stay observable

Stream live state, contain failures, and report measured per-run and cumulative savings.

No four-worker ceiling

Every machine gets its own plan.

The scheduler does not encode a fixed “4×” target. Four lanes belong only to the controlled public benchmark.

64runtime worker ceiling

Actual concurrency is usually lower and is recalculated from current conditions.

CPU-boundStarts from physical cores; understands Apple performance/efficiency topology and SMT.

I/O-boundAllows bounded oversubscription when the host has safe capacity.

Shared hostCoordinates concurrent Codex tasks instead of letting each one assume it owns the computer.

Under pressureBacks off for live CPU load, memory pressure, thermal state, and battery policy.

Tailored execution

Useful across the work
developers actually wait for.

One safety core, with workload-specific planning instead of a single generic “run in parallel” switch.

TEST

Large test suites

Use a bounded native worker pool when case independence, fixtures, ordering, and shared resources permit it.

BUILD

Build and package graphs

Fan out independent packages or targets while preserving caches, artifacts, and success gates.

CONTAINERS

Docker workflows

Plan multiple images and services inside the detected Docker Desktop CPU and memory envelope.

DATA / PYTHON

Research and batch jobs

Separate preparation, validation, figures, and deterministic merges; advise on safe program-level refactors.

Public evidence

Proof you can inspect.

Release checks, platform evidence, and long-running benchmark artifacts are bound to the published source commit—not copied from a marketing spreadsheet.

macOSVerified

4.00×

Controlled four-lane overlap benchmark.

5m 10swall time

15m 30ssaved this run

4h 39mcumulative saved

WindowsEvidence pending

—

Controlled four-lane overlap benchmark.

Pendingwall time

Pendingsaved this run

Pendingcumulative saved

What “~4×” means here: the controlled benchmark deliberately runs four equal, independent lanes. It verifies overlap, lifecycle, and accounting. It is not AtomLane’s maximum concurrency or a universal speedup claim.

Full machine-readable report ↗ macOS benchmark artifact ↗ Windows benchmark artifact ↗ GitHub Actions ↗ commit 7bb687545e · generated 2026-10-12T10:56:13+00:00

Install in two commands

Give Codex an execution lane.

Register the public marketplace, install AtomLane, then open a new Codex task so its Hook and tools load cleanly.

中文文档 ↗Latest release ↗

codex plugin marketplace add cloudguo123/atomlane

codex plugin add atomlane@atomlane

Use $accelerate-local-work to inspect this project and run only the work proven safe in parallel. Keep progress visible and report measured time saved.

macOS is Stable. Native Windows remains a scoped Preview; review the Windows contract before production use.

Run more.Break less.

Parallelism is compiled,not guessed.