Agentic Coding Flywheel · Jeffrey Emanuel · as of October 5, 2026
The Agentic Coding Flywheel puts 85% of the effort before the first line of code
Its reasoning: even a 6,000-line plan still fits in a model’s head, and the code it describes never will. So the thinking happens in the plan, argued over by several AI models, and a swarm of agents builds what it describes.
This guide is written by EnsoDynamics, which makes the Enso Method. We compare our method with others as fairly as we can, and we recommend another method when it fits you better.
Two philosophies at a glance
Agentic Coding Flywheel
Do the hard thinking while the whole plan still fits in a model’s head, polish it across models until it converges, then let a swarm of interchangeable agents build it.
The Enso Method
Define the work in full too, then keep those documents as the system’s source of truth and build on proposals the people who decide confirm.
Where they differ
- Specs serve one change Specs live as long as the system Agentic Coding Flywheel sits toward “Specs serve one change”; the Enso Method sits at the “Specs live as long as the system” end.
- The developer at the keyboard decides Someone who isn’t at the keyboard decides Agentic Coding Flywheel sits at the “The developer at the keyboard decides” end; the Enso Method sits at the “Someone who isn’t at the keyboard decides” end.
- Stops and asks Proposes an answer to be confirmed, keeps building Agentic Coding Flywheel sits in the middle; the Enso Method sits at the “Proposes an answer to be confirmed, keeps building” end.
- Each task starts fresh Lessons carried to later work Agentic Coding Flywheel sits at the “Lessons carried to later work” end; the Enso Method sits toward “Lessons carried to later work”.
What they share
- One change at a time A body of work defined in full up front Agentic Coding Flywheel sits at the “A body of work defined in full up front” end; the Enso Method sits at the “A body of work defined in full up front” end.
- Starts from a prompt Intent written down first Agentic Coding Flywheel sits at the “Intent written down first” end; the Enso Method sits at the “Intent written down first” end.
Neither end of a line is the right one. Every comparison uses the same lines, so you can compare across methods.
Think while the whole system still fits in context
“The models are far smarter when reasoning about a plan that is very detailed and fleshed out but still trivially small enough to easily fit within their context window.” That line from Jeffrey Emanuel is the root of the method he calls the Agentic Coding Flywheel, and the reason it asks for “85%+ of your time, attention, and energy” to go into planning.
Once a plan turns into code, the system grows too big for any model to see whole. Planning tokens are also far cheaper than implementation tokens, so you can afford round after round of revision. Hence its update to the woodworking maxim: “Check your beads N times, implement once.”
The flywheel is what happens next. Every agent session is logged and mined, and what worked goes back into the prompts, tools and instructions, so each project starts from a better baseline than the last.
Rival models write the plan; a swarm builds it
- Plan. You explain the goals and user workflows. GPT Pro, Claude, Gemini and Grok each draft a plan; one merges them into a “best of all worlds” hybrid, and fresh rounds refine it until the changes turn small, often at 3,000 to 6,000 lines.
- Beads. The plan becomes hundreds of beads: self-contained tasks carrying their dependencies, reasoning and required tests, polished four to six times or more.
- Swarm. Five to fifteen interchangeable agents (Claude Code, Codex, Antigravity) share one checkout. A graph tool picks each one’s next bead; Agent Mail carries claims and file reservations.
- Tend. Every 10 to 30 minutes you look for stuck beads, revive confused agents and send review prompts.
- Harden. Agents review their own and each other’s code until reviews come back clean, and write tests “without using mocks.”
Its author: “Once you have the beads in good shape based on a great markdown plan, I almost view the project as a foregone conclusion at that point.”
Built for one person running many agents
The Flywheel suits a builder who owns the product, wants something large and new built quickly, and will spend hours on a plan first: about three for one complex feature in its own example. As its author puts it, “my team is a bunch of agents.”
It asks for real setup: a rented Linux server with 48 to 64 GB of memory, a stack of command-line tools its installer puts in place, and agents running with their permission prompts switched off. Its site puts the cost at $440 to $656 a month, the server plus Claude Max and ChatGPT Pro, though the guide says you can start with one provider.
It suits you less for small changes (“obviously overkill”), or when a person must read each change before it lands.
Where the Enso Method made a different call
Both define the work in full before any code, write it all down, distrust tests built on mocks, and run several agents in one checkout, each forbidden to touch the others’ changes. They part on three things.
What the plan becomes once the build starts
Agentic Coding Flywheel
The beads take over: “once you’re in ‘bead space’ you never look back at the markdown plan.” The plan stays as a record; the next feature gets a new one.
The Enso Method
The PRD and technical design stay the source of truth for the life of the system, amended before every change.
Choose by situation: to build fast from one great plan, the Flywheel is lighter; for a system others will maintain, the Enso Method keeps a current account of what it should do.
How a plan gets better
Agentic Coding Flywheel
Rival models critique it, asked for “new features, changed features,” and you choose what survives.
The Enso Method
Each open question is researched against the real code and data, and refinement “never adds scope.” Business questions go to their owners as proposals while the build continues.
Choose by situation: if the product is yours and you want it better, many models help; if someone else decided what to build, research what is true and keep to it.
Who reads the code before it lands
Agentic Coding Flywheel
Agents do: “There is no pull request, no human reviewer, no approval gate.” They review their own and each other’s work, then push to the main branch.
The Enso Method
Two fresh-eyes agent passes check the work and harden its tests; then, by default, you read the diff and make the commit.
Choose by situation: for speed across many agents, the Flywheel; where a person answers for every change, the Enso Method.
Where Agentic Coding Flywheel shines
- Plans tested against several minds. Four frontier models draft and critique each plan, and a prompt insisting it missed “at least 80” items keeps the reviewer digging. The Enso Method’s refinement uses one model.
- A swarm that shrugs off crashes. Any agent can take any bead, so losing one costs “some slowdown and some wasted tokens.” The Enso Method has no layer for coordinating many agents.
- Memory that compounds. Every session is indexed, and rules distilled from them fade unless reinforced. The Enso Method keeps no store of lessons.
- Guards, not just instructions. After an agent wiped uncommitted work, a hook began blocking destructive commands outright. The Enso Method relies on written rules.
How the Enso Method differs, and why
The hard thinking comes before any code in both methods; they part over what happens after. The Agentic Coding Flywheel turns its plan into self-contained tasks for a swarm of agents that review each other’s work, and starts each new feature from a fresh plan. The Enso Method keeps its requirements as the system’s source of truth, amended before every change, sends business questions to their owners as proposals, and by default has a person read each change before it lands.
Get started
The Agentic Coding Flywheel is free, under an MIT licence with a rider that excludes OpenAI and Anthropic. The method is set out at agent-flywheel.com/complete-guide, and the setup wizard is at agent-flywheel.com.