A December 2022 firsthand look at building an early agent harness for GPT-3/ChatGPT
OwariDa · x · 2026-11-18
Joel Eriksson revisits the first agent harness he built on Dec 7, 2022, days after ChatGPT launched, for GPT-3/ChatGPT-3.5:
- Tool calling: implemented by working around the fact that LLMs then weren't trained to support it natively;
- Architecture: subagents plus an agent loop that ran autonomously until an objective was reached;
- Context bottleneck: handled with progressive summarization and overlapping chunkwise processing of subgoal data.
His contemporaneous tweets (screenshot) also predicted key agent directions: training LLMs to emit special sequences for interacting with other systems, "self-guided training" (given an objective, the model identifies what to train on next, then loops), and optimizing LLMs for dialogue as a prefrontal-cortex-like executive function. Over three years later, these read as early blueprints for today's agent and continual-learning research.
More from coding & agent
- Founder vibe-codes a dashboard to track all his AI agents across machines — Ben_Reinhardt · 2026-11-18
- A ChatGPT subscription isn't AI-native: build skilled agents, tools access and review loops — kylegawley · 2026-11-18
- NVIDIA AI Podcast returns with Hugging Face on open models and agent safety — nvidia · 2026-11-18
- Building a remote MCP server with write tools and OAuth: hard-won lessons from a solo dev — Exciting_Shake4017 · 2026-11-18
- Crossmint open-sources Menta, a plugin that lets AI agents complete purchases at 91% success across 3,000 stores — heyshrutimishra · 2026-11-18
- Qodo launches Auto Approval Workflows to fix the coding agent PR bottleneck — heyshrutimishra · 2026-11-18