Agent & AI Insights
Claude Code know-how, heterogeneous agents, harness engineering, and more. What the Marblo team measured running these, and what we learned adopting them in-house.
- AI Agents
Measured
Your Cost Table Is Empty and Both the Parser and the Router Are Correct
Wiring a new harness into per-agent cost tracking has three attachment points, not two. The third fails silently. Here is the diagnostic fingerprint for a bug you cannot find by reading the code, plus how unregistered model IDs invent charges.
- AI Agents
Measured
How Do You Know an Agent Is Working? PTY Output Is Wrong in Both Directions
Treat terminal bytes as proof of work and finished agents look busy while working agents look idle. The load average hit 44 on a 12-core machine with 11 zombies while the cleanup tool reported nothing to reclaim. Define turns by boundaries, not noise.
- AI Agents
Marblo Build Notes 1 — The Closed Loop Release
From the closed loop to Orchebrain, Web Tabs, and risk-based independent testers. Eight new features in this Marblo release, plus a preview of Marblo CI.
- AI Agents
Marblo Build Notes 2 — Completing a Mission with the Closed Loop
The October 2026 hypemarc.com improvements: how a mission progressed from orchestrator reviews and ticket assignments to worker PRs, independent testing, merge gates, and production deployment confirmation.
- AI Agents
Marblo Build Notes 3 — Watching Work with Agent Mosaic
Agent Mosaic brings working agents' terminals onto one screen in real time. How teams, leaders, and developers can use it to follow what several AI agents are doing.
- AI Agents
Marblo Build Notes 4 — Following Mission Progress with Orchebrain
Orchebrain groups scattered tickets into missions and shows progress through a summary card and development, testing, and review dots. How teams can use its autonomous progress limits and intake rate controls.
- AI Agents
Marblo Build Notes 7 — Distributing Work Across Multiple AI Accounts
Marblo distributes work across Claude, GPT, and other accounts by remaining allowance. In the hypemarc.com work, a Claude account fell below the reserve threshold and a new task went to a Codex account.
- AI Agents
Marblo Build Notes 8 — Independent Testers for Important Changes
A separate AI tester verifies important payment, permission, and deployment changes, including counterexamples. The hypemarc.com work shows independent security handoffs, orchestrator reviews, and a build note PR returned for corrections.
- AI Agents
Marblo Build Notes 9 — Setting Direction through Telegram
Give instructions to the orchestrator through Telegram and receive progress and problem-cause reports. In the hypemarc.com work, the CEO's direction led to updated examples, build notes, version labels, and a deployment report.
- AI Agents
AI Agent Orchestration Platforms in 2026 — LangGraph, CrewAI, AutoGen, and Marblo Compared
An engineering-grade comparison of the major AI agent orchestration platforms in 2026. Where each one shines, where each one breaks, and which choice fits which workload — from prototype to multi-team production.
- AI Agents
Codex 5.5 GOAL Mode: Agents That Work From a Goal
OpenAI Codex's latest GOAL mode changes how autonomous agents are used. We analyze the shift from imperative commands to goal-driven execution, and how it compares to Marblo's natural language orchestrator.
- AI Agents
Why Heterogeneous AI Agents Beat a Single Model: Claude, Codex, and Grok on One Board
Why we recommend heterogeneous agent orchestration over single-vendor setups. The quality and cost benefits of splitting reasoning, generation, and verification across models by strength.
- AI Agents
Claude Code Subagents vs. Real Multi-Agent Orchestration: What's the Difference?
We dissect the gap between Claude Code's 'subagent' pattern and genuine heterogeneous multi-agent orchestration. Single-model N agents vs. heterogeneous N agents, CLI vs. kanban board, shared context vs. physical isolation.