Technology / WORLD

Cohere study compares coding agent team strategies

In Cohere’s ProgramBench tests, planning-focused teams reached similar scores at lower reported cost.

Reading sizeSources ↓Share story ↗
Conceptual illustration of technology, processors and connected information.Illustration
Symbolic illustration

Cohere researchers published a controlled study comparing ways to organize coding-agent teams. On 200 ProgramBench tasks, an asynchronous GPT-5.6 Sol team delegating only planning reported 63.6% reward at a cost of $249; a GPT-5.6 Sol team delegating all stages scored 63.2% at $510. Both figures were averages across three runs.

A GPT-5.6 Sol lead coordinating 16 GPT-5.6 Luna planning workers scored 61.8% at $155, compared with 63.6% at $249 for the all-Sol planning team. This fixed 16-worker configuration was run once. Cohere says the findings are specific to its test harness; some model comparisons used a single trial, and the results do not establish general agent performance.

Data

Timeline

Cohere study compares coding agent team strategies

Date and scope of the development reported in the primary source.

  1. · Occurred

    Cohere study compares coding agent team strategies

    In Cohere’s ProgramBench tests, planning-focused teams reached similar scores at lower reported cost.

Methodology

The announcement date published by the source is recorded as one event.

Updated:

Table

Cohere ProgramBench team results

ProgramBench results from Cohere’s coding-agent team evaluation; reward and inference cost are reported as stated by the source.

GPT-5.6 Sol; async lead delegates planning only63.6249200 tasks; mean of 3 runs
GPT-5.6 Sol; async lead delegates all stages63.2510200 tasks; mean of 3 runs
GPT-5.6 Sol lead; 16 GPT-5.6 Luna planning workers61.8155Fixed 16-worker configuration; 1 run
Methodology

Reward is the mean share of hidden behavioral tests passed across 200 ProgramBench tasks. The first two asynchronous GPT-5.6 Sol configurations average three runs; the fixed configuration with 16 GPT-5.6 Luna workers was run once. Results are from Cohere’s own test harness.

Updated:

Structured coverage

Subjects and places

Subject proposals

Geographic references

Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.

  • Canada · Place role: Coverage · Place type: Country or area · Location precision: Regional
    Evidence
    • Explicit preserved editorial country/region; no city or national centroid inferredcohere.com ↗
Taxonomy version: IPTC Media Topics 2026-Q2

Continue reading

Technology

ProgramGym builds 1,648 verifiable training environments for coding agents

The October 9 arXiv preprint describes environments for reproducing programs from scratch and a staged agent workflow.

arXiv preprint by researchers affiliated with Xiamen University, Washington University in St. Louis, the Chinese Academy of Sciences and Unisound1 min read
Data included