Social·Community
Project Charters as a Bridge to Agentic Harnesses: A Non-ML Path to Multi-Agent Workflows
A practitioner with a non-traditional ML background describes a multi-model workflow spanning GPT, Claude, Perplexity, and Notion, and argues project charters — documents defining objectives, constraints, agent roles, source-of-truth locations, state, handoff expectations, and completion criteria — serve as a bridge to automated agentic harnesses. The proposed path: write the architecture down first, run the loop manually, then automate the patterns that prove stable. The framing matters for teams building multi-agent coordination without deep infrastructure experience.
10 engagement1 sourceSat, Oct 3, 2026, 04:32 PM
AI Overview
- A practitioner outlines a multi-model workflow spanning GPT, Claude, Perplexity, and Notion, with Notion serving as a persistent second-brain.
- Project charters define objective, constraints, agent roles, source-of-truth locations, state, handoff expectations, and completion criteria before automation.
- He proposes the progression charter → explicit workflow → repeatable handoffs → stable patterns → automation → harness, with the human initially acting as the loop.
- He argues handoff quality, shared context, role boundaries, and stopping conditions can matter as much as individual model capabilities.
- He recommends Sam Bhagwat's Principles of Building AI Agents and Patterns for Building AI Agents (with Michelle Gienow), plus IBM Technology and Google explainers.
Related
Social·Community
Builder says biggest multi-agent architecture win was making agents replaceable, not central
A practitioner reworked a production multi-agent system so that state, authority, and evidence live outside the agents, making workers swappable across models, scripts, or humans. Completion is verified by reading back external state, and reasoning is demoted to handling only genuine uncertainty. The post frames agent architecture as organizational design, arguing the durable value is the surrounding system, not the LLM.
A practitioner reworked a production multi-agent system so that state, authority, and evidence live outside the agents, making workers swappable across models, scripts, or humans. Completion is verified by reading back external state, and reasoning is demoted to handling only genuine uncertainty. The post frames agent architecture as organizational design, arguing the durable value is the surrounding system, not the LLM.
6 engagement1 sourceredditSun, Oct 4, 2026, 06:52 PM
Products·building
Shared state tool for multi-agent AI coding sessions
A tool the author built after failing to manage context across multiple AI agents (writer, reviewer, deployer) via manual Markdown handoff notes. It gives agents shared verified state — what changed, what's blocked, and evidence a task is done — rather than just conversation history. Aimed at developers running multi-agent coding workflows.
A tool the author built after failing to manage context across multiple AI agents (writer, reviewer, deployer) via manual Markdown handoff notes. It gives agents shared verified state — what changed, what's blocked, and evidence a task is done — rather than just conversation history. Aimed at developers running multi-agent coding workflows.
34 engagement1 sourceredditTue, Sep 1, 2026, 08:45 AM
Products·building
AI Project Manager for coding agents
Nexus is a tool that acts as an AI project manager for coding agents, maintaining architectural consistency as projects grow. It addresses the problem of agents forgetting design decisions and introducing bugs by providing a structured context. Built with just 2 prompts, it improves coding agent performance more than switching LLMs.
Nexus is a tool that acts as an AI project manager for coding agents, maintaining architectural consistency as projects grow. It addresses the problem of agents forgetting design decisions and introducing bugs by providing a structured context. Built with just 2 prompts, it improves coding agent performance more than switching LLMs.
7 engagement1 sourceredditSun, Aug 2, 2026, 11:57 AM
Social·Community
Reddit user builds AGENTS.md that turns Astra into a Codex multi-agent orchestrator across Luna, Terra, and Sol
A developer built a project-specific AGENTS.md that turns Astra into the orchestrator of a Codex multi-agent coding workflow, delegating to Luna, Terra, and Sol instead of the user manually shuttling handoffs between models. The setup separates model choice from reasoning effort, minimizes token usage during delegation, and requires Astra to report and learn from its own routing decisions after each task. The author keeps it project-specific pending further testing and notes the Codex desktop UI lacks clear live views of model and effort per subagent.
A developer built a project-specific AGENTS.md that turns Astra into the orchestrator of a Codex multi-agent coding workflow, delegating to Luna, Terra, and Sol instead of the user manually shuttling handoffs between models. The setup separates model choice from reasoning effort, minimizes token usage during delegation, and requires Astra to report and learn from its own routing decisions after each task. The author keeps it project-specific pending further testing and notes the Codex desktop UI lacks clear live views of model and effort per subagent.
56 engagement1 sourceredditFri, Sep 11, 2026, 07:06 PM
Products·building
Neo Roundtable — a shared conversation that gives multiple AI tools the same project context
Neo Roundtable lets users give ChatGPT, Claude, and coding agents access to one shared conversation so context and decisions carry across tools. It targets developers and knowledge workers who use several AI assistants on the same project and currently have to re-explain context in each chat. LLMs are the consumers of the shared context, keeping decisions consistent across tools.
Neo Roundtable lets users give ChatGPT, Claude, and coding agents access to one shared conversation so context and decisions carry across tools. It targets developers and knowledge workers who use several AI assistants on the same project and currently have to re-explain context in each chat. LLMs are the consumers of the shared context, keeping decisions consistent across tools.
0 engagement1 sourceredditFri, Sep 25, 2026, 06:40 AM