How to configure thinking_level / reasoning effort for Antigravity Managed Agent in Interactions API?

Hi team,

I am testing the remote Antigravity agent harness in the Gemini Interactions API using the Python SDK (google-genai), (antigravity-preview-09-2026)

1. What is the default reasoning/thinking level for the orchestrator model inside antigravity-preview-09-2026 when running in environment="remote"? Is it hardcoded to medium or perhaps it is adaptive?

2. How can I configure the thinking level for the agent?

interaction = client.interactions.create(
    agent="antigravity-preview-09-2026",
    environment="remote",
    agent_config={
        "type": "antigravity",
        "model": "gemini-4-argon",
        # Is there a field like thinking_level or reasoning_effort supported here?
    },
    input="...",
)

Sorry I just found nothing in the documentation on this topic
Thanks in advance

The reason you could not find this in the documentation is that thinking_level is not exposed in agent_config. Here is the exact architectural breakdown from the upstream Interactions API specification:

1. Default Reasoning Behaviour in antigravity-preview-09-2026

When invoking agent="antigravity-preview-09-2026" in a remote sandbox, reasoning is adaptive and dynamic.

The orchestrator manages its internal reasoning budget server-side across multi-step tool interactions (code generation, terminal commands, workspace inspection) rather than adhering to a static thinking quota.

2. Upstream Schema: AntigravityAgentConfig

In the Gemini v1beta Interactions API specification, the AntigravityAgentConfig schema accepts only three properties:

Schema: AntigravityAgentConfig
  properties:
    type: string (required, "antigravity")
    model: string (optional, e.g. "gemini-4-argon" or "gemini-3.8-flash")
    max_total_tokens: string (int64, optional)

Fields like thinking_level or reasoning_effort do not exist within agent_config and will either be silently disregarded or trigger an HTTP 400 invalid_argument error.


How to Control Reasoning & Spend

Depending on your objective, you have two supported levers:

Option A: Constrain the Agent via max_total_tokens

If your goal is to prevent runaway execution or cap token usage during complex tasks, pass max_total_tokens in agent_config:

interaction = client.interactions.create(
    agent="antigravity-preview-09-2026",
    environment="remote",
    agent_config={
        "type": "antigravity",
        "model": "gemini-4-argon",
        "max_total_tokens": "32768",  # int64 string
    },
    input="Analyse the repository and produce an architectural summary.",
)

Option B: Direct Model Interactions (When Explicit thinking_level is Mandatory)

If your pipeline strictly requires explicit control over reasoning depth ("low", "medium", or "high"), use direct model interactions rather than the managed agent wrapper, passing thinking_level inside generation_config:

interaction = client.interactions.create(
    model="gemini-3.8-flash",
    input="Analyse the repository and produce an architectural summary.",
    generation_config={
        "thinking_level": "low",      # "low" | "medium" | "high"
        "thinking_summaries": "auto", # "auto" | "none"
    },
)

For managed agents, if you need to rein in speculative planning without switching to raw model interactions, the most effective technique is steering the agent via prompt directives (e.g. instructing it to execute directly without lengthy speculative design phases).

hi thanks for contributing: +1 @dllhell 's answer