Integrationv2.4

AutoGen + Meridian

Wire Microsoft AutoGen multi-agent groups directly into Meridian's runtime. Agents share a single configuration surface — tune models, tool policies, and conversation limits from one env block.

Back to Docs

Configuration Reference

10 variables

Variable	Type	Default	Description
`AUTOGEN_MODEL`	string	`gpt-4o`	Primary model for agent conversations. Supports any OpenAI-compatible endpoint.
`AUTOGEN_MAX_ROUNDS`	number	`12`	Maximum conversation turns before the orchestrator forces termination.
`AUTOGEN_TEMPERATURE`	number	`0.7`	Sampling temperature passed to every agent in the group chat.
`AUTOGEN_ENABLE_TOOLS`	boolean	`true`	Allow agents to invoke registered Python functions during reasoning.
`AUTOGEN_LOG_LEVEL`	string	`INFO`	Python logging level for the AutoGen runtime. One of DEBUG, INFO, WARNING, ERROR.
`AUTOGEN_CODE_EXEC_TIMEOUT`	number	`30`	Seconds before a code-execution cell is killed by the sandbox.
`AUTOGEN_RETRY_ATTEMPTS`	number	`3`	How many times an agent retries a failed tool call before surfacing the error.
`AUTOGEN_SEED`	number	`42`	Random seed for reproducible agent runs. Set to 0 for non-deterministic.
`AUTOGEN_CACHE_SEED`	boolean	`true`	Enable LLM response caching keyed on (prompt, seed). Requires disk space.
`AUTOGEN_HUMAN_INPUT_MODE`	string	`NEVER`	When to pause for human feedback. Values: ALWAYS, TERMINATE, NEVER.

AUTOGEN_MODELstring

default: gpt-4o

Primary model for agent conversations. Supports any OpenAI-compatible endpoint.

AUTOGEN_MAX_ROUNDSnumber

default: 12

Maximum conversation turns before the orchestrator forces termination.

AUTOGEN_TEMPERATUREnumber

default: 0.7

Sampling temperature passed to every agent in the group chat.

AUTOGEN_ENABLE_TOOLSboolean

default: true

Allow agents to invoke registered Python functions during reasoning.

AUTOGEN_LOG_LEVELstring

default: INFO

Python logging level for the AutoGen runtime. One of DEBUG, INFO, WARNING, ERROR.

AUTOGEN_CODE_EXEC_TIMEOUTnumber

default: 30

Seconds before a code-execution cell is killed by the sandbox.

AUTOGEN_RETRY_ATTEMPTSnumber

default: 3

How many times an agent retries a failed tool call before surfacing the error.

AUTOGEN_SEEDnumber

default: 42

Random seed for reproducible agent runs. Set to 0 for non-deterministic.

AUTOGEN_CACHE_SEEDboolean

default: true

Enable LLM response caching keyed on (prompt, seed). Requires disk space.

AUTOGEN_HUMAN_INPUT_MODEstring

default: NEVER

When to pause for human feedback. Values: ALWAYS, TERMINATE, NEVER.

Quickstart

.env

# Primary model for agent conversations. Supports any OpenAI-compatible endpoint.
AUTOGEN_MODEL=gpt-4o

# Maximum conversation turns before the orchestrator forces termination.
AUTOGEN_MAX_ROUNDS=12

# Sampling temperature passed to every agent in the group chat.
AUTOGEN_TEMPERATURE=0.7

# Allow agents to invoke registered Python functions during reasoning.
AUTOGEN_ENABLE_TOOLS=true

# Python logging level for the AutoGen runtime. One of DEBUG, INFO, WARNING, ERROR.
AUTOGEN_LOG_LEVEL=INFO

# Seconds before a code-execution cell is killed by the sandbox.
AUTOGEN_CODE_EXEC_TIMEOUT=30

# How many times an agent retries a failed tool call before surfacing the error.
AUTOGEN_RETRY_ATTEMPTS=3

# Random seed for reproducible agent runs. Set to 0 for non-deterministic.
AUTOGEN_SEED=42

# Enable LLM response caching keyed on (prompt, seed). Requires disk space.
AUTOGEN_CACHE_SEED=true

# When to pause for human feedback. Values: ALWAYS, TERMINATE, NEVER.
AUTOGEN_HUMAN_INPUT_MODE=NEVER