@dabit3: Since you don't need a frontier model typing every line of code, we can optimize subagents beyond just cost. /lightning…

X AI KOLs Following Tools

Summary

Lightning Orchestrator is an agent skill that uses fast subagents (SWE-1.7 Lightning) for implementation, achieving ~5x faster execution while keeping a frontier model as planner and reviewer. It works with Devin and supports parallel execution for larger tasks.

Since you don't need a frontier model typing every line of code, we can optimize subagents beyond just cost. /lightning skill offloads implementation to 1,000 tok/s subagents, resulting in ~5x faster execution. it keeps frontier (or default) as the planner + reviewer, hands implementation to SWE-1.7 Lightning from @cognition. Cool thing about this pattern is that on larger or fanned-out work, it compounds! Works with @DevinAI CLI + Desktop. https://github.com/dabit3/lightning-orchestrator…
Original Article
View Cached Full Text

Cached at: 07/15/26, 07:58 PM

Since you don’t need a frontier model typing every line of code, we can optimize subagents beyond just cost.

/lightning skill offloads implementation to 1,000 tok/s subagents, resulting in ~5x faster execution.

it keeps frontier (or default) as the planner + reviewer, hands implementation to SWE-1.7 Lightning from @cognition.

Cool thing about this pattern is that on larger or fanned-out work, it compounds!

Works with @DevinAI CLI + Desktop. https://github.com/dabit3/lightning-orchestrator…


dabit3/lightning-orchestrator

Source: https://github.com/dabit3/lightning-orchestrator

Lightning

Fast execution, smart orchestration for Devin.

Lightning is an agent skill that pairs the intelligence of your active Devin model with the speed and cost efficiency of SWE-1.7 Lightning. The active model plans, makes decisions, and reviews the result, while a dedicated Lightning subagent performs the implementation.

Why Lightning?

Using one model for every step can be unnecessarily slow or expensive. Lightning separates the work into two roles:

RoleResponsibility
Active Devin modelUnderstands the request, resolves important decisions, creates the work order, and reviews the result
lightning-executorInspects the code, implements the change, runs focused checks, and reports evidence

This provides a fast implementation path without giving up higher-level reasoning or independent review.

How it works

When you invoke the skill, the orchestrator:

  1. Frames the objective, acceptance criteria, constraints, and validation needs.
  2. Inspects enough of the repository to resolve material product or architecture questions.
  3. Sends a self-contained work order to the lightning-executor subagent.
  4. Reviews the resulting diff and verification evidence against the original request.
  5. Applies trivial fixes directly or resumes the same executor with targeted feedback when a correction is needed.
  6. Reports the completed changes, verification results, and any remaining risks.

The skill uses one executor by default to minimize context duplication and cost. For clear, low-risk tasks the preflight reads are issued in the same round as the dispatch, review reads are batched the same way, and long-running suites run in the background during diff review — keeping orchestrator overhead off the critical path. Purely explanatory or read-only requests are handled directly when delegation would not add value.

Three further optimizations keep handoff overhead in check:

  • Trivial edits skip delegation. Every handoff has a roughly fixed coordination cost — the work order and report are each written by one side and read by the other — so a few clearly understood, low-risk lines are edited directly instead of being routed through an executor.
  • Follow-up work resumes the same executor. Resuming keeps the executor’s accumulated repository context and prompt cache warm, so later work orders skip rediscovery instead of re-paying a cold start.
  • Long exploratory tasks get steering checkpoints. Instead of one open-ended order, the work is split into milestone-scoped resumes with a review between each, so unproductive directions are caught early rather than after a long run.

When fanning out to parallel executors, the orchestrator embeds its already-established shared context (build commands, conventions, key paths) into every work order, since parallel workers cannot see each other’s discoveries and would otherwise repeat the same research.

Requirements

  • Devin with support for project skills and custom subagent profiles
  • Access to the swe-1.7-lightning model

Installation

Install both the skill and its executor profile. They must be installed together because the skill explicitly delegates implementation to lightning-executor.

Skills CLI

Install the skill for Devin in the current project:

npx skills add dabit3/lightning-orchestrator --agent devin

The Skills CLI installs SKILL.md packages, but not Devin-specific custom agent profiles. Install the required executor separately:

mkdir -p .devin/agents/lightning-executor
curl -fsSL \
  https://raw.githubusercontent.com/dabit3/lightning-orchestrator/main/.devin/agents/lightning-executor/AGENT.md \
  -o .devin/agents/lightning-executor/AGENT.md

For a global installation, add --global and place the executor in Devin’s global configuration directory:

npx skills add dabit3/lightning-orchestrator --agent devin --global
mkdir -p ~/.config/devin/agents/lightning-executor
curl -fsSL \
  https://raw.githubusercontent.com/dabit3/lightning-orchestrator/main/.devin/agents/lightning-executor/AGENT.md \
  -o ~/.config/devin/agents/lightning-executor/AGENT.md

Manual installation

If you already cloned this repository, copy both directories into your project:

mkdir -p .devin/skills .devin/agents
cp -R /path/to/lightning-orchestrator/.devin/skills/lightning .devin/skills/
cp -R /path/to/lightning-orchestrator/.devin/agents/lightning-executor .devin/agents/

For a manual global installation, copy them into your user configuration directory:

mkdir -p ~/.config/devin/skills ~/.config/devin/agents
cp -R /path/to/lightning-orchestrator/.devin/skills/lightning ~/.config/devin/skills/
cp -R /path/to/lightning-orchestrator/.devin/agents/lightning-executor ~/.config/devin/agents/

Start a new Devin session after installation so the skill and agent profile are discovered.

Usage

Invoke Lightning with a concrete software-engineering task:

/lightning add pagination to the users endpoint and update its tests
/lightning reproduce and fix the checkout total rounding bug
/lightning refactor the cache adapter without changing its public API

Good requests describe the observable outcome and any important constraints. Lightning discovers routine implementation details and local repository conventions on its own.

Behavior and safeguards

Lightning is designed to:

  • prioritize correctness and safety over speed or cost
  • preserve existing user changes
  • implement the smallest coherent diff
  • follow repository-specific instructions and conventions
  • add or update focused tests when appropriate
  • verify the result before reporting success
  • avoid destructive operations and external side effects without explicit approval
  • avoid unrelated refactors, dependencies, generated files, and documentation

If the custom executor profile is unavailable, Lightning stops and reports the missing profile rather than silently switching to another model.

Permissions

Both definitions pre-approve the read-only inspection commands on the critical path — git status, git diff, and git log — so fresh sessions do not stall on approval prompts just to inspect repository state. Writes, edits, and all other commands follow your normal approval flow.

To remove per-edit approval prompts entirely — and to let background fan-out executors edit files without a foreground resume — opt in by adding edit to the executor’s allowed permissions in .devin/agents/lightning-executor/AGENT.md:

permissions:
  allow:
    - Exec(git status)
    - Exec(git diff)
    - Exec(git log)
    - edit

Only enable this if you are comfortable with the executor editing files without prompting.

Repository structure

.devin/
├── agents/
│   └── lightning-executor/
│       └── AGENT.md
└── skills/
    └── lightning/
        └── SKILL.md
  • .devin/skills/lightning/SKILL.md defines the orchestration workflow.
  • .devin/agents/lightning-executor/AGENT.md defines the model-pinned implementation agent.

Customization

You can adapt the workflow by editing the two definitions:

  • Change orchestration policy, review requirements, or delegation criteria in SKILL.md.
  • Change the executor model, tool restrictions, implementation protocol, or permissions in AGENT.md.

Keep the skill’s referenced profile name and the executor’s name field aligned if you rename lightning-executor.

Similar Articles

@dabit3: 1,000 tok/s vs 85 tok/s visualized

X AI KOLs Timeline

Nader Dabit visualizes the speed difference between 1,000 tok/s subagents and 85 tok/s, highlighting that lightning skill offload enables ~5x faster execution by using subagents for implementation while keeping frontier models as planners and reviewers.