Agent settings

Everything OmniCoreAgent(agent_config={...}) takes. Each description is the comment above the setting in the code.

Setting	Type	Default	What it does
`agent_version`	`str \| None`	`None`	Recorded as the trace’s agent_version; a content hash of the harness (prompt, tools, model, and settings) is used when it is not set.
`request_limit`	`int`	`0`	Most model calls (LLM requests) one run may make; over it, the run ends with termination_reason resource_limit. 0 is no limit. max_steps is usually the limit you want; for dollars, or limits across runs, use governance budgets.
`total_tokens_limit`	`int`	`0`	Most tokens (in and out) one run may use; over it, the run ends with termination_reason resource_limit. 0 is no limit.
`max_steps`	`int`	`50`	Most turns of the agent loop one run may take (a turn is a model call and the tools it asked for); then it ends with termination_reason max_steps. 1 to 1000.
`tool_call_timeout`	`int`	`180`	Seconds one tool call may take before it fails with a timeout. 2 to 1000.
`subagent_timeout`	`int \| None`	`None`	Seconds a delegation (spawn_subagents) may take. None: a worker is bounded by its own step cap and the run’s deadline, not by the timeout for one tool call.
`mcp_enabled`	`bool`	`False`	Accepted for compatibility with 0.3.x and has no effect: MCP is on when mcp_tools are given.
`enable_advanced_tool_use`	`bool`	`False`	For large tool sets: a tools_retriever tool the model uses to find tools by describing what it needs; the matching tools are offered on the next turn.
`enable_subagents`	`bool`	`False`	Let the agent spawn focused workers (spawn_subagents), each under the same policy and budgets with its own linked trace. Turns workspace files on.
`worker_profiles`	`list[dict[str, Any]]`	`[]`	The kinds of worker the lead may spawn, each with its own model, reasoning effort, tools, steps and deny/ask rules; the lead picks one per worker by its description. Empty: workers are as the lead, less spawning. Needs enable_subagents. See core/worker_profiles.py.
`enable_agent_skills`	`bool`	`False`	Offer the skills found in skills_dir: their instructions, files and scripts.
`skills_dir`	`str \| None`	`None`	Where skills are found; None is ./.agents/skills in the working directory. A harness names its own, so the task’s directory stays the task’s.
`completion_review`	`int`	`0`	How many times a final answer is reviewed before it is accepted: the model is asked for each requirement and the check that showed it, and the run goes on in the same trace. Off by default.
`skill_script_env`	`list[str]`	`[]`	Environment variables host skill scripts receive beyond the minimal set (PATH, HOME, locale, TMPDIR, TERM); secrets are not passed by default.
`run_lease_seconds`	`int`	`60`	A live run refreshes its heartbeat within this many seconds; a run whose heartbeat is older can be recovered by another process.
`run_retention_days`	`int \| None`	`30`	Finished run records (completed, failed, blocked, cancelled, timeout) are removed from the memory store this many days after the run started; None keeps every record forever. A run still waiting for a person or a resume is never removed.
`code_mode`	`dict[str, Any]`	`{}`	Code mode: a run_code tool that runs Python in Monty (omnicoreagent[codemode]).
`agents_md`	`dict[str, Any]`	`{}`	A project’s own instructions for the agent (AGENTS.md), by path.
`memory_config`	`dict[str, Any]`	see below	How much of a session’s earlier history a new run is given. mode “sliding_window” keeps the last `value` messages; “token_budget” keeps the latest messages that fit in `value` tokens. The default (10,000 messages) gives the whole history in practice, and context_management keeps each model call under its token budget. summary.enabled replaces what falls out of the window with a summary; summary.retention_policy is whether the summarized messages stay in the store (“keep”) or are deleted (“delete”).
`enable_workspace_files`	`bool`	`True`	File tools over the agent’s workspace (ls, read_file, write_file, edit_file, glob, grep, …). The workspace itself is set by workspace_config.
`guardrail_config`	`dict[str, Any]`	`{}`	Tunes the prompt-injection guardrail’s detection: strict_mode, sensitivity, allowlist_patterns, blocklist_patterns, and the rest of DetectionConfig. Empty uses its defaults.
`guardrail_mode`	`str`	`"full"`	Where the prompt-injection guardrail screens: “full” (the request and every tool result), “input_only” (the request), or “off”.
`privacy_config`	`dict[str, Any]`	see below	Personal data (email, phone, SSN, card numbers) redacted per boundary. By default only the record (telemetry and its exports) is redacted; the model, memory, files, stream and answer see the real data. redact_public covers everything returned to the application: run()‘s answer and errors, and OmniServe’s responses.
`context_management`	`dict[str, Any]`	see below	Keeps what one model call is sent under a budget, within a run: in mode “token_budget”, `value` is tokens (100,000); in “sliding_window”, it is messages. Past threshold_percent of it, older messages are truncated or summarized (strategy), keeping the preserve_recent latest. memory_config decides the history a run starts from; this keeps the run itself in bounds as it grows.
`tool_offload`	`dict[str, Any]`	see below	A tool result over the thresholds is saved to the workspace, and the model gets a preview and an artifact it can read in full.
`governance_config`	`dict[str, Any]`	see below	The policy (a profile, a policy, or a policy file), budgets, the sandbox provider and its manifest, and how unanswered approvals are handled. On by default with the permissive-dev profile; {“enabled”: False} turns it off, and is refused alongside a policy, profile or budgets it would drop. Enabling it without a profile means interactive-dev.
`workspace_config`	`WorkspaceConfig \| dict[str, Any] \| None`	`None`	Where the agent’s workspace lives: workspace_dir on local disk (default ./workspace), or S3 / R2 storage.

Defaults in full

`memory_config`

How much of a session’s earlier history a new run is given. mode “sliding_window” keeps the last value messages; “token_budget” keeps the latest messages that fit in value tokens. The default (10,000 messages) gives the whole history in practice, and context_management keeps each model call under its token budget. summary.enabled replaces what falls out of the window with a summary; summary.retention_policy is whether the summarized messages stay in the store (“keep”) or are deleted (“delete”).

{
  "mode": "sliding_window",
  "summary": {
    "enabled": false,
    "retention_policy": "keep"
  },
  "value": 10000
}

`privacy_config`

Personal data (email, phone, SSN, card numbers) redacted per boundary. By default only the record (telemetry and its exports) is redacted; the model, memory, files, stream and answer see the real data. redact_public covers everything returned to the application: run()‘s answer and errors, and OmniServe’s responses.

{
  "categories": [
    "credit_card",
    "email",
    "phone",
    "ssn"
  ],
  "enabled": true,
  "redact_memory": false,
  "redact_model_io": false,
  "redact_public": false,
  "redact_stream": false,
  "redact_telemetry": true,
  "redact_workspace": false
}

`context_management`

Keeps what one model call is sent under a budget, within a run: in mode “token_budget”, value is tokens (100,000); in “sliding_window”, it is messages. Past threshold_percent of it, older messages are truncated or summarized (strategy), keeping the preserve_recent latest. memory_config decides the history a run starts from; this keeps the run itself in bounds as it grows.

{
  "enabled": true,
  "mode": "token_budget",
  "preserve_recent": 4,
  "strategy": "truncate",
  "threshold_percent": 75,
  "value": 100000
}

`tool_offload`

A tool result over the thresholds is saved to the workspace, and the model gets a preview and an artifact it can read in full.

{
  "enabled": true,
  "max_preview_lines": 10,
  "max_preview_tokens": 150,
  "threshold_bytes": 2000,
  "threshold_tokens": 500
}

`governance_config`

The policy (a profile, a policy, or a policy file), budgets, the sandbox provider and its manifest, and how unanswered approvals are handled. On by default with the permissive-dev profile; {“enabled”: False} turns it off, and is refused alongside a policy, profile or budgets it would drop. Enabling it without a profile means interactive-dev.

{
  "allow_static_high_risk_approvals": false,
  "allow_test_sandbox_runtime": false,
  "approval_mode": "suspend",
  "approval_resolver": null,
  "budgets": null,
  "enabled": true,
  "policy": null,
  "policy_path": null,
  "profile": "permissive-dev",
  "project_root": null,
  "sandbox_config": null,
  "sandbox_manifest": null,
  "sandbox_runtime": null,
  "workspace_bridge": null
}

Get Started

Make It Safe

Run It

See and Improve

Build

How It Works

Reference

HTTP API (OmniServe)

Releases

Agent settings

Agent settings

Defaults in full

`memory_config`

`privacy_config`

`context_management`

`tool_offload`

`governance_config`

​Agent settings

​Defaults in full

​memory_config

​privacy_config

​context_management

​tool_offload

​governance_config

Agent settings

Defaults in full

`memory_config`

`privacy_config`

`context_management`

`tool_offload`

`governance_config`