Matthew Berman on Relay: Routing Coding Work Across Models

SUMMARY

In 30 seconds

Matthew Berman presents Relay, a skill in the Forward Future GitHub organization, and argues that combining frontier and lower-cost models can speed up coding tasks while reducing expense. He promotes the project as free and open source; this is his own project rather than an independent product review. His example compares a mixed-model workflow with single-model costs, but does not establish that every task will be faster or produce equally good code. The [Relay repository](https://github.com/Forward-Future/relay) describes a model-agnostic planning, execution and review workflow and is MIT-licensed.

1Worker agentsBerman shows multiple worker agents and says the example uses GPT-5.6 Terra rather than GPT-5.6 Sol for the work.
2Three stagesHe describes a frontier model planning and architecting, lower-cost workhorse models writing code, and another frontier model reviewing it.
3One cost exampleA mixed route reportedly cost $25.55 compared with $81 for Fable alone; the displayed comparison also lists $46.50 for GPT-5.6 alone. Results depend on the task and configuration.
4Free releaseBerman says he is giving Relay away for free and that it is open source. The linked repository belongs to Forward Future, which he promotes in the description.

Takeaways

Relay routes planning, execution and review to models selected for different roles.
Berman's example uses Fable to plan, Composer 2.5 to execute and GPT-5.6 Sol to review.
His displayed $81 versus $25.55 comparison is a particular example, not a general price or quality guarantee.
Berman promotes Relay as free and open source, and the linked repository is under the Forward Future organization.

TopicsAgent orchestration Coding agents AI cost & performance Agent skills Code review & verification

FULL TEXT

Read the edited text by chapter

Show transcript

Transcript only

Delegating coding work

Berman shows worker agents and describes using a frontier model for planning and cheaper models for implementation. A separate frontier model reviews the code.

Here you can see Relay has kicked off multiple worker agents. In this example, they use GPT-5.6 Terra as a workhorse rather than GPT-5.6 Sol. Relay lets a frontier model such as Fable plan and architect the task, then delegates the code writing to workhorse models that I say are less expensive and faster.

Another frontier model then reviews the code. My aim is to get code that is just as good while paying much less than using a frontier model for every step. That is the intended benefit, not a demonstrated guarantee for every task; the result still needs review and testing.

Cost example and release

He gives a mixed-model cost example and says the task finished faster. He offers Relay as a free, open-source skill.

SLIDE 1 / 1Example cost comparisonThe on-screen figures compare Fable alone, GPT-5.6 alone and a mixed workflow for this example; they are not universal prices.

In one example, we used Fable for planning, Composer 2.5 for execution and GPT-5.6 Sol for review. I say this took the cost from $81 to just over $25 and completed the task faster. The on-screen comparison lists $81 for Fable alone, $46.50 for GPT-5.6 alone and $25.55 for the mixture. These are figures from this example, not promised prices or a controlled proof of equal quality.

I am giving Relay away for free and it is open source. The description links to the Forward Future repository; its README describes the planning, execution and independent review workflow and lists an MIT license. This is a promotion of my own offering.