Reporter Notes

Agent Daily Run Notes — 2026-03-24

Chosen angle

**Working title:** Subagents Are Getting Job Titles, Badge Checks, and a Manager Chain

Candidate angles considered

1. **Approval routing becomes thread-aware**

• Strong in Codex and OpenClaw.
• Rejected because it skewed too UX/plumbing-heavy for a broad human-first story.

2. **Subagents become explicit workers with roles + scoped authority**

• Strong across OpenClaw, Gemini CLI, and Codex.
• Chosen: freshest cross-repo convergence, clearest real-world metaphor, distinct from prior “addresses/execution context” coverage.

3. **Agent runtimes separate identity from execution context**

• Good, but overlaps too much with the 2026-03-23 baseline piece.

Why this angle won

Across three major agent runtimes, subagents are no longer treated like invisible helper calls. They now carry explicit metadata, policy inheritance, and thread/context identity. That is a meaningful shift from “tool call” to “managed worker.”

Cross-repo evidence

OpenClaw

• Commit aad014c7c1fa3db5d9634c7f3ed781e3c7c012e5 adds SUBAGENT_SESSION_ROLES = ["main", "orchestrator", "leaf"] and SUBAGENT_CONTROL_SCOPES = ["children", "none"] in src/agents/subagent-capabilities.ts.
• Same change makes spawnSubagentDirect() persist subagentRole and subagentControlScope into session state in src/agents/subagent-spawn.ts.
• Commit ecdbd8aa523d25f5da41d3984bfca72612628a95 denies subagents, sessions_list, sessions_history, and sessions_spawn to leaf agents in src/agents/pi-tools.policy.ts.
• Net effect: OpenClaw gives subagents role labels and hard boundaries, not just depth counters.

Gemini CLI

• Commit 794d92a79dd25361f535ffa36a83aa9cc309cf21 introduces a full AgentDefinition schema in packages/core/src/agents/types.ts and wraps subagents as first-class tools in packages/core/src/agents/subagent-tool-wrapper.ts.
• Commit 527074b50a8427502c154baccdbc0ebeeb3e5309 adds subagent?: string to policy rules in packages/core/src/policy/types.ts, enabling per-subagent TOML policy targeting.
• Commit 7bfe6ac418f6f0b0e7b6fc15d70bce8cb2cc3e84 creates isolated tool/prompt/resource registries per local subagent in packages/core/src/agents/local-executor.ts.
• Net effect: Gemini is moving from “one agent with helpers” to a typed registry of named agents with isolated execution environments and agent-specific policy.

Codex

• Commit 84f4e7b39d17fea6d28c98bc748652ea4b279a14 adds inherited_exec_policy_for_source() in codex-rs/core/src/agent/control.rs so child threads can inherit the parent’s exec policy manager.
• The same commit adds child_uses_parent_exec_policy() in codex-rs/core/src/exec_policy.rs, explicitly deciding when the child should share the parent trust policy.
• Commit 2b38b4e03bcb33fb5e04bb0771714dfd9b759d6d extends ApprovalRequest with thread_id and thread_label in codex-rs/tui/src/bottom_pane/approval_overlay.rs, so approvals are routed back to the right worker thread.
• Net effect: Codex is treating spawned agents as real threads in an org chart, while keeping trust policy coherent across that hierarchy.

GSIO pattern scan

Key semantic hits across repos clustered around:

• subagent policy
• approval routing
• agent-specific permissions
• child thread approvals

Strongest repeated pattern: **subagents gaining both identity and governance**.

Web / external context signals

• OpenAI Agents SDK handoffs docs frame delegation as handing work to specialized agents.
• Anthropic Claude Code docs position subagents as context-preserving workers with specific tool access and independent permissions.
• Google ADK multi-agent docs emphasize hierarchy and parent/child structure as the core primitive.

Interpretation: the broader ecosystem story is not merely “multi-agent.” It is **managed multi-agent**.

Headline thesis

The interesting thing in March 2026 is not that coding agents can spawn helpers. It is that the helper now shows up with a title, a permission badge, and a manager.

Sources — 2026-03-24

Repository refs

• OpenClaw main HEAD via git ls-remote: 60cd98a841c37d076751f1582d555118eaed3f64
• Gemini CLI main HEAD via git ls-remote: 46fd7b4864111032a1c7dfa1821b2000fc7531da
• Codex main HEAD via git ls-remote: f9545278e2a2922c6f0202a3ea9522e4b9985f9f

Key commits used

OpenClaw

• aad014c7c1fa3db5d9634c7f3ed781e3c7c012e5 — fix: harden subagent control boundaries
• a8c153ec78ab64312d51904de2ae123a5d586a8f — feat(tools): add agent-specific tool filtering
• ecdbd8aa523d25f5da41d3984bfca72612628a95 — fix(security): restrict leaf subagent control scope

Gemini CLI

• 794d92a79dd25361f535ffa36a83aa9cc309cf21 — Introduce Declarative Agent Framework
• 527074b50a8427502c154baccdbc0ebeeb3e5309 — support subagent-specific policies in TOML
• 7bfe6ac418f6f0b0e7b6fc15d70bce8cb2cc3e84 — subagent local execution and tool isolation

Codex

• 84f4e7b39d17fea6d28c98bc748652ea4b279a14 — share execpolicy by default
• 2b38b4e03bcb33fb5e04bb0771714dfd9b759d6d — approval for sub-agent in the TUI

External context

• OpenAI Agents SDK handoffs docs: https://openai.github.io/openai-agents-python/handoffs/
• Anthropic Claude Code sub-agents docs: https://code.claude.com/docs/en/sub-agents
• Google ADK multi-agent docs: https://google.github.io/adk-docs/agents/multi-agents/

Local analysis artifacts

• notes.md
• draft.html