Batch 10 Notes — Multi-agent orchestrators (Codex-native + cross-runtime)

Frameworks covered

Total files: 4 + (9 × 11) = 103 files + this notes file

Taxonomy findings

New archetypes surfaced in this batch

PR lifecycle platform (agent-orchestrator-composio): background daemon + issue tracker polling + git-worktree isolation + CI/review reaction system + plugin slot architecture. Most complete end-to-end automated coding pipeline in the corpus. Reaction system (ci-failed → send-to-agent, changes-requested → send-to-agent) is a unique primitive.

Cross-runtime orchestration harness (maestro-orchestrate): single canonical src/ tree compiled to 4 AI coding tools (Gemini CLI, Claude Code, Codex, Qwen Code). src-first, generated-runtime is a unique architectural approach. Server-side design gate and DAG validation add hard enforcement not seen in other harnesses.

Linux-kernel module system (amplifier): ultra-thin kernel (~2,600 lines) + protocol contracts (Tool/Provider/Orchestrator/ContextManager/Hook) + git-installable bundle marketplace. Research demonstrator from Microsoft MADE:Explorations. Not production-ready but conceptually distinct.

Adversarial QA system (aistack): dedicated adversarial agent that actively tries to break code using 6 attack vectors (SQL injection, XSS, SSRF, auth bypass, path traversal, deserialization) with VERDICT: APPROVE/REJECT gate. JWT/RBAC agent identity tokens are unique in the corpus.

Python DAG DSL (agentflow-berabuddies): >> operator for DAG construction, fanout(), merge(), on_failure >> back-edges, native EC2/ECS/SSH remote execution. Jinja2 template chaining for cross-node output injection.

Patterns appearing across multiple frameworks in this batch

1. Reaction/event-driven self-repair: AO (ci-failed/changes-requested), symphony-openai (webhook events) — autonomous CI loop without human polling
2. tmux PTY isolation: AO (runtime-tmux), CAO (tmux per agent session) — both use tmux as process isolation layer
3. Multiple agent backends: AO (Claude Code/Codex/Aider/OpenCode), CAO (7 CLIs), maestro (4 runtimes) — portability as explicit design goal
4. Hard-gated approval: maestro (design gate, code review gate), symphony-openai (approval_policy config), AO (escalateAfter timeout) — different mechanisms but same pattern
5. MCP as coordination backbone: CAO (handoff/assign/send_message MCP primitives), maestro (17-tool MCP server), amplifier (MCP as primary tool protocol)

Cross-batch observations

symphony-openai vs symphony-sincover: same "symphony" name, completely unrelated projects. OpenAI's Codex automation system vs a Roo Code custom-modes pack by sincover. Name collision confirmed — treat as separate taxonomy entries.

ruflo = claude-flow rename: confirmed via README and package structure. No architectural divergence, only rebranding + plugin count growth (22 → 33). Delta-only treatment is correct.

CAO (AWS Labs) vs AO (ComposioHQ): closest structural peers in entire corpus. Both: background daemon, tmux PTY, dashboard, multi-agent-backend. CAO differentiates: MCP handoff protocol, cron scheduling, event plugins (Discord/Slack/Telegram). AO differentiates: native PR lifecycle + tracker integration, reaction system, 7-slot plugin architecture, openclaw-plugin for Claude Code IDE control.

amplifier stars (3,095) vs maturity mismatch: Microsoft research demo with high visibility but documented as not production-ready. Bundle marketplace concept (git-installable modules with protocol contracts) is the most sophisticated extensibility model in the corpus, more so than AO's 7-slot plugin system or claude-flow's 33 plugins.

Data quality notes

• symphony-openai: Elixir as primary language is unexpected for an AI coding agent system; Codex runs in JS App Server mode as subprocess — Elixir is the orchestration layer, not the agent runtime
• agentflow-berabuddies: Pi node (routes to 10+ providers via Pi AI API) is documented but not tested — treat as declared capability
• aistack: JWT/RBAC agent identity tokens not seen in any other framework — unique primitive, could be seed-worthy for enterprise compliance angle
• maestro: COBOL/HLASM/z/OS mainframe agent coverage is unique in the entire corpus — explicitly targets IBM legacy enterprise stack
• opencode-orchestrator: Korean-language AGENTS.md (evidence-based anti-hallucination) — primary documentation not in English; treat evidence-based promise as declared, not independently verified

Recommended follow-up

• aistack JWT/RBAC pattern warrants a dedicated primitive entry in seed-comparison.md
• maestro's src-first generated-runtime approach warrants comparison with any other multi-runtime frameworks found in other batches
• amplifier bundle marketplace: check if any batch covered similar git-installable module ecosystems