WIP: Multi-agent support: Agent, Protocol, MultiAgentEnv#965

Draft

nph4rd wants to merge 25 commits intoPrimeIntellect-ai:mainfrom

nph4rd:multiagent-no-opponent-conditioning

Commits on Mar 21, 2026

add MultiAgentEnv for turn-based multi-agent environments
nph4rd
committed
rename Actor to Agent, add Protocol abstraction
nph4rd
committed
require Protocol in MultiAgentEnv, simplify docstrings
nph4rd
committed
update docstrings
nph4rd
committed
add multi-agent reward functions for heterogeneous rewards
nph4rd
committed
compute per-agent advantages for multi-agent rewards
nph4rd
committed
include all rewards in per-agent rewards for multi-agent training
nph4rd
committed
add opponent-conditioned baselines for multi-agent advantage estimation
nph4rd
committed
add debug logging for opponent-conditioned baselines
nph4rd
committed
add trajectory structure debug
nph4rd
committed
debug extras and state keys
nph4rd
committed
remove opponent-conditioned baselines for comparison test
nph4rd
committed
add per-agent baselines for multi-agent advantage computation
nph4rd
committed
debug: log multiagent func detection
nph4rd
committed
fix score_rollout to support multi-agent reward functions
nph4rd
committed
add diagnostic logging to score_rollout multiagent path
nph4rd
committed
add game state logging to score_rollout diagnostics
nph4rd
committed
add trajectory length and completion flags to diagnostics
nph4rd
committed
log error details in multiagent rollout and score_rollout diagnostics
nph4rd
committed
normalize messages from build_agent_prompt before storing in trajectory
nph4rd
committed
remove diagnostic logging
nph4rd
committed
add per-agent reward metrics for multi-agent environments
nph4rd
committed
add per-agent model routing for multi-policy lora training
nph4rd
committed
point textarena to fork with kuhn poker fixes
nph4rd
committed

Commits on Mar 30, 2026

fix get_free_port_pair macos compatibility with zmq bind
nph4rd
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

WIP: Multi-agent support: Agent, Protocol, MultiAgentEnv#965

WIP: Multi-agent support: Agent, Protocol, MultiAgentEnv#965
nph4rd wants to merge 25 commits intoPrimeIntellect-ai:mainfrom
nph4rd:multiagent-no-opponent-conditioning

Commits on Mar 21, 2026

add MultiAgentEnv for turn-based multi-agent environments

rename Actor to Agent, add Protocol abstraction

require Protocol in MultiAgentEnv, simplify docstrings

update docstrings

add multi-agent reward functions for heterogeneous rewards

compute per-agent advantages for multi-agent rewards

include all rewards in per-agent rewards for multi-agent training

add opponent-conditioned baselines for multi-agent advantage estimation

add debug logging for opponent-conditioned baselines

add trajectory structure debug

debug extras and state keys

remove opponent-conditioned baselines for comparison test

add per-agent baselines for multi-agent advantage computation

debug: log multiagent func detection

fix score_rollout to support multi-agent reward functions

add diagnostic logging to score_rollout multiagent path

add game state logging to score_rollout diagnostics

add trajectory length and completion flags to diagnostics

log error details in multiagent rollout and score_rollout diagnostics

normalize messages from build_agent_prompt before storing in trajectory

remove diagnostic logging

add per-agent reward metrics for multi-agent environments

add per-agent model routing for multi-policy lora training

point textarena to fork with kuhn poker fixes

Commits on Mar 30, 2026

fix get_free_port_pair macos compatibility with zmq bind