Breaking

Bello Packages Coding Models Into a Supervised Agent Pipeline and Reports Lower Codex Usage

October 3, 2026 at 10:16 EDT

Software Dev & Coding
AI Agents
Security

Bello 0.7.1, released on September 30, 2026, packages Codex, Claude Code and API-based models into a supervised coding pipeline instead of treating one model as the whole engineering loop. The single-maintainer Python project reports that its Efficient Budget setup used 66.3% less of a weekly Codex limit than Raw GPT-5.6 Sol XHigh across matched ProgramBench runs while scoring 1.45% higher, with a much longer average runtime.

Sep 30, 2026 · Bello

Coding models, split into roles

Bello wraps a coding agent with supervision, independent review, adversarial testing and local log filtering so a run is managed like a small engineering pipeline.

66.3%

less weekly Codex limit in Efficient Budget

+1.45%

higher mean score in the same comparison

26.41%

relative gain from Sol Ultra C+A

599 MB

local log-distiller model download

Weekly limit use, drawn to scale

15.63%

Raw Codex

5.27%

Bello Budget

The budget profile used about one-third as much weekly limit as the raw baseline.

Pipeline shape

Coder

builds in a disposable workspace

→

Supervisor

watches drift and risky actions

→

Reviewer

checks evidence and completeness

→

Adversary

tries to break the result

What improves

Lower measured quota use, independent review, adversarial probing and a final report that records changed files, checks and remaining risks.

What slows down

The budget run averaged 1:49:44 versus 28:27 for raw Codex, and some cost-saving modes showed small score drops.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.