Feature proposal: inherited recursion-depth budget for Agent.as_tool delegation
メンテナーはふだん 1 日以内に返信
まだ誰も着手していません。
評価
- 難易度
- 5/5
- 見積もり時間
- 1週間以上
- 初心者へのやさしさ
- 35/100
調査の方向性
Start with Agent.as_tool() in src/agents/agent.py, then trace how nested calls create Runner.run() and Runner.run_streamed() executions and how their turn limits are resolved. Review the existing runtime governance discussion #2775 and per-agent max_turns request #551 before deciding whether the proposed inherited depth limit fits the SDK. Done means maintainers settle the scope and semantics, including concurrent task inheritance and resumed workflows.
索引モデルが issue の本文から書いたものです。
説明
Summary
I would like to propose an opt-in, inherited delegation-depth limit for Agent.as_tool() to help applications bound recursive or multi-agent delegation trees.
This is not a claim of a security vulnerability or a defect in max_turns. I understand that max_turns limits an individual Runner.run(), and a nested agent tool starts another run. The motivation is to offer an explicit SDK-side safety primitive when applications require a depth limit over a delegation chain, rather than just each run.
Reproduced behavior (no API access)
I ran a fully offline agents.testing.ScriptedModel test against public main commit cd26bbbc52847845fde73b9952b51e62e2bcbdc0 (v0.24.0 source) using distinct call IDs and a deliberately finite, four-delegation script.
- Outer
Runner.run(..., max_turns=1) - Recursive
Agent.as_tool(..., max_turns=1) - Unmodified SDK: 5 model calls occur before the outer run raises
MaxTurnsExceeded. This is consistent with per-run limits. - Candidate guard with
max_agent_tool_depth=1: 2 model calls occur; the next nested agent tool fails before starting another child run, including through streaming.
The source boundary is Agent.as_tool() in src/agents/agent.py, where each nested invocation creates its own Runner.run() / Runner.run_streamed() with a separately resolved turn limit.
Proposed narrow, compatible implementation
- Append optional
max_agent_tool_depth: int | None = NonetoAgent.as_tool().Nonepreserves current behavior. - Track active agent-tool depth in an async context using
ContextVar, withfinallyreset. - Propagate a configured limit through other agents'
as_tool()wrappers, even if those descendants do not configure a limit. Children may tighten but cannot relax an inherited bound. - Reject excessive delegation before another child run is started.
- No dependencies, persisted-state changes or wire-protocol changes.
Reproducible public research repository: https://github.com/BlackRonin-coder/openai-agents-recursive-budget-research
It contains the unmodified SDK reproducer, opt-in two-file patch, synthetic tests, and local verification log. The published candidate passed 1,188 relevant tests, 63 skipped, on macOS with Python 3.12.15; Ruff checks and a pristine-source patch-application check passed. I am not claiming comprehensive CI, independent validation or OpenAI endorsement.
Limitations and maintainer questions
This is a nesting-depth limit, not a root-owned cumulative model-turn, token, cost, wall-clock or cross-process execution budget. Application-initiated independent Runner.run() calls are outside this implementation. Further review is needed around concurrent task inheritance and resumed/persisted workflows.
- Would an opt-in inherited agent-tool depth limit be useful as a first-class SDK control?
- Or would you prefer a root-owned, cumulative execution budget with explicit checkpoint/cancellation semantics?
- If there is an existing supported mechanism addressing this exact delegation-depth use case, I would welcome a pointer and will revise or close the proposal.
Related but broader prior art: runtime governance discussions (e.g. #2775), and the existing per-agent max_turns feature request (#551). I did not find a directly equivalent recursive delegation-depth control in the currently indexed issues/PRs. This is offered as independently reproducible engineering research, not a claim of being the first person to explore recursive agent budgets.
Thank you for considering it.
- 主要言語
- Python
- スター
- 30.4k
- フォーク
- 4.9k
- 平均マージ
- 15時間 42分
- マージ済み PR（30日）
- 124
環境構築
- Dockerfile・Docker Compose ファイルなし
- プルリクエストのテンプレートあり
- コントリビューションガイドを読む
はじめの一歩
- issue を最後まで読み、次にプロジェクトのコントリビューションガイドを読みます。
- 着手することを issue にコメントします — 二人が同じ作業をするのを防げます。
- リポジトリをフォークし、ブランチを切って変更します。
- issue 番号を参照したプルリクエストを送ります。
openai/openai-agents-python のほかの issue
-
難易度 2/5 1〜3時間 初心者へのやさしさ 77/100
openai/openai-agents-python#5451 · コメント 1 件 ·
メンテナーはふだん 1 日以内に返信
-
fix(typing): explicitly specialize ToolContext[Any] in ToolFunctionWithToolContext再び着手できるかも @AbarnaaSree が 11 日前に担当しましたが、オープン中のプルリクエストはありません。 オープン
難易度 1/5 1時間未満 初心者へのやさしさ 85/100
openai/openai-agents-python#5387 · コメント 4 件 ·
メンテナーはふだん 1 日以内に返信
-
難易度 2/5 1〜3時間 初心者へのやさしさ 53/100
openai/openai-agents-python#5450 ·
メンテナーはふだん 1 日以内に返信
-
function_tool rejects valid arguments when a parameter uses Pydantic strict mode (Enum, date, UUID)対応中かも @Lahari25 が 1 日前に担当しました。 オープン
難易度 3/5 1〜2日 初心者へのやさしさ 65/100
openai/openai-agents-python#5448 · コメント 2 件 ·
メンテナーはふだん 1 日以内に返信
-
難易度 3/5 1〜2日 初心者へのやさしさ 65/100
openai/openai-agents-python#5447 · コメント 3 件 ·
メンテナーはふだん 1 日以内に返信
openai/openai-agents-python の issue をすべて見る
似ている issue
-
priority middle
難易度 1/5 1時間未満 初心者へのやさしさ 62/100
KATO-Hiro/AtCoderClans#12940 ·
メンテナーはふだん 1 日以内に返信
-
難易度 1/5 1時間未満 初心者へのやさしさ 76/100
huggingface/huggingface_hub#5283 ·
メンテナーはふだん 1 日以内に返信
-
難易度 1/5 1時間未満 初心者へのやさしさ 78/100
メンテナーはふだん 1 日以内に返信
-
enhancement
難易度 1/5 1時間未満 初心者へのやさしさ 86/100
メンテナーはふだん 1 日以内に返信
-
bug example
難易度 2/5 1〜3時間 初心者へのやさしさ 86/100
メンテナーはふだん 1 日以内に返信