Project Charters as a Bridge to Agentic Harnesses

Telos_in_the_Void · reddit · 2026-10-04

A practitioner from a non-traditional ML/engineering background shares how he built a multi-model workflow (GPT, Claude, Perplexity, Notion as a persistent second brain) and why the core question shifted from "which model is best" to continuity, role ownership, source of truth, handoffs, task state, and completion criteria.

His answer is project charters: defining objective, constraints, context, agent roles, source-of-truth locations, current state, handoff expectations, and completion criteria.

He proposes an evolution path:

> Charter → explicit workflow → repeatable handoffs → stable patterns → automation → harness

The human can initially be the loop. You do not need orchestration software first — write the architecture down, then observe where handoffs fail, context gets lost, and responsibilities blur; those are the pieces worth turning into code.

He cautions that adding more agents does not automatically create a better system: handoff quality, shared context, role boundaries, and stopping conditions can matter as much as individual model capability.

Resources shared: IBM Technology's short explainers on agent skills, memory, MCP and context engineering; Google's developer content; and Sam Bhagwat's Principles of Building AI Agents and Patterns for Building AI Agents.

Original post →

More from coding & agent

coding & agent channel →