Clankers are very smart, but…

Agents are great at thinking and planning, but not at long-running execution.

Gantry breaks work down into agent-sized chunks. No LLM in the loop at the top level, every agent gets a fresh context.

Create a plan, any way you like. Gantry will read it and…

(This is the default pattern. It is not hardcoded: write your own.)

Build

Gantry sizes up your plan. If it's a big one, it gets broken into milestones. A test command is written and run once before any work starts, and again when every milestone is done.

gantry plan.md runs build.toml

plan plan.md

milestones write-milestones.md test command write-project-gate.md

test baseline

for each milestone milestone.toml

test final

Milestone

Each milestone is split into sprints. First an environment brief and a milestone test command are written; every agent in the milestone reads the brief. When the last sprint is done, an agent reviews the whole milestone.

milestone.toml runs once per milestone

for each milestone

brief write-environment.md sprints write-sprints.md test command write-milestone-gate.md

test baseline

for each sprint sprint.toml review review.md

Sprint

A fresh agent implements the sprint. Gantry runs the tests. Another fresh agent reviews the diff. Gantry runs the tests again.

sprint.toml runs once per sprint

for each sprint

execute execute.md

test Gantry runs it

review review.md

test Gantry runs it

Executor execute.md

Implements the sprint brief. If the brief cannot be done as written, says so and stops instead of improvising.
Reviewer review.md

Reviews the executor's diff against the brief. Applies safe fixes; never widens the scope.
when the tests fail

tests ✗ red
troubleshoot troubleshoot.md
test Gantry runs it

retry fresh sprint, up to attempt_ceiling

STOP handover.md
Troubleshooter troubleshoot.md

Repairs a red gate. Ends in one of three states: repaired, retry with a note for a fresh attempt, or handover — a report for the operator, and the run stops.

Merge

Every run takes place in its own git worktree, on its own branch. Gantry does the git work itself, with no LLM: it merges the home branch's tip into the run branch, runs the tests, merges the run branch back, and runs the tests once more. An agent is launched only for a conflict.

worktree run branch

integrate home branch tip

test project gate

merge into the home branch

test on the home branch

when the home branch has moved during the run

conflict textual or tests ✗
resolve resolve-merge.md
test project gate

merge only if green
Conflict Resolver resolve-merge.md

Resolves textual conflicts and test failures the merge caused, inside the worktree. If the result cannot be made green, Gantry stops; the home branch is untouched.

Until

The Gantry version of a "loop": the whole machinery above, repeated until a goal is reached.

gantry --until goal.md runs until.toml

goal check

build test plan review merge fix

goal check

build test plan review merge fix

goal green

Map

A list of tasks, built one after another. No milestones, no sprints: one agent per task, then the tests. A failed task is recorded and the run goes on.

gantry --map tasks.md runs map.toml

source tasks.md

roster write-map-roster.md

for each task map_task

implement the task's prompt

test the task's gate

Roster Writer write-map-roster.md

Turns the task source into a roster: one task file per task, each with its own prompt and test command.
when the tests fail

tests ✗ red

repair fresh agent

test Gantry runs it

failed recorded

next task the run goes on
Repairer

A second agent with the same task prompt, launched once when the task's tests fail.

2026-08-25
Gantry Patterns: A language for agent orchestration

To make Gantry more generally useful, I created an agent orchestration data-format. It's not pretty, but it works.
2026-08-06
The Gantry Tapes: Building an agent sequencer in public

I'm building Gantry in public by posting stats, case studies and video replays of the Gantry runs that implemented hundreds of Gantry features.
2026-08-02
What has Gantry actually been used to build?

Gantry itself and this website, a Godot colony sim, about a dozen Rails apps, and the sibling TUI projects Skriv and Brev.

Gantry Patterns: A language for agent orchestration

The Gantry Tapes: Building an agent sequencer in public

What has Gantry actually been used to build?