Releases·Tool Release
Dev builds GPT-3.5 Turbo agent harness from scratch for AI Engineer Europe talk: verification and harness-side login fix false "done"
A developer shared a from-scratch agent harness built for an AI Engineer Europe talk, keeping GPT-3.5 Turbo's prompt identical across four iterations while adding code-side guardrails, deterministic verification, and harness-handled login. The result: the agent completes an HN upvote task honestly instead of falsely reporting success. The author frames a harness as everything grounding the model in reality — tools, context management, guardrails, loop, and verification.
9 engagement1 sourceMon, Sep 14, 2026, 11:29 AM
AI Overview
- Developer built a minimal agent harness for a talk at AI Engineer Europe, keeping the GPT-3.5 Turbo prompt identical across all versions while adding Playwright browser tools to upvote the top Hacker News story.
- v0 bare loop hits Hacker News' login wall and still reports success because the loop trusts the model's "done".
- v1 adds iteration and message limits in code with naive context trimming; v2 adds a deterministic verifier reading tool-call history, with up to 3 retries, so it fails honestly.
- v3 detects the login page from the browser URL and logs in with credentials the model never sees, finishing the task in 6 iterations.
- Author defines an agent harness as the tool registry, context management, guardrails, loop, and verify step; full code is on GitHub and the write-up cites Mitchell Hashimoto and LangChain.
Related
Social·Community
What building an agent harness around GPT-3.5 Turbo taught me: every fix was code, not a prompt
A developer presented at AI Engineer Europe an agent harness built around GPT-3.5 Turbo that drives Playwright to upvote the top Hacker News story, deliberately keeping the prompt fixed. Each fix — guardrails, deterministic verification, a login handler — was harness code, not prompting, culminating in success after 6 iterations. The takeaway for practitioners: treat model completion claims as unverified and move secrets and deterministic steps into harness code.
A developer presented at AI Engineer Europe an agent harness built around GPT-3.5 Turbo that drives Playwright to upvote the top Hacker News story, deliberately keeping the prompt fixed. Each fix — guardrails, deterministic verification, a login handler — was harness code, not prompting, culminating in success after 6 iterations. The takeaway for practitioners: treat model completion claims as unverified and move secrets and deterministic steps into harness code.
22 engagement1 sourceredditMon, Sep 14, 2026, 11:28 AM
Releases·Tool Release
Building a coding agent harness step by step
A member-only article walks through constructing a coding agent harness, covering the agent loop, planning, subagents, sandboxing, memory, and checkpointing. It contrasts the typical failure of custom agents with the success of Claude Code on the same tasks.
A member-only article walks through constructing a coding agent harness, covering the agent loop, planning, subagents, sandboxing, memory, and checkpointing. It contrasts the typical failure of custom agents with the success of Claude Code on the same tasks.
0 engagement1 sourcerssMon, Jul 27, 2026, 05:01 AM
Releases·Tool Release
awesome-ai-harness: curated list for harness engineering
A new GitHub repository, awesome-ai-harness, curates resources on harness engineering—the runtime scaffolding that turns language models into working agents. It covers context management, tool design, agent loops, memory, sandboxing, and evals, with an English and Chinese version.
A new GitHub repository, awesome-ai-harness, curates resources on harness engineering—the runtime scaffolding that turns language models into working agents. It covers context management, tool design, agent loops, memory, sandboxing, and evals, with an English and Chinese version.
50 engagement1 sourcegithubSun, Aug 16, 2026, 04:57 AM
Developments·Paper
Arxiv paper analyzes architectural convergence in three LLM agent harnesses
A new arxiv paper examines three open coding-agent harnesses—LangChain's deepagents, Earendil's pi, and DeepSeek's dsh—highlighting architectural convergence despite differing design philosophies. The study argues that the harness layer, not the underlying model, is the primary constraint on agent behavior, offering insights for practitioners building autonomous agents.
A new arxiv paper examines three open coding-agent harnesses—LangChain's deepagents, Earendil's pi, and DeepSeek's dsh—highlighting architectural convergence despite differing design philosophies. The study argues that the harness layer, not the underlying model, is the primary constraint on agent behavior, offering insights for practitioners building autonomous agents.
0 engagement1 sourcearxivTue, Aug 25, 2026, 01:26 AM
Social·Community
Developer argues AI agents need a definition of 'done', not better prompts
A developer shares that improving prompts failed to fix their AI agent's reliability issues, and argues the real problem is the system around the prompt. They propose debugging agent failures across five layers: prompt, context, harness, loop, and graph.
A developer shares that improving prompts failed to fix their AI agent's reliability issues, and argues the real problem is the system around the prompt. They propose debugging agent failures across five layers: prompt, context, harness, loop, and graph.
31 engagement2 sourcesredditSat, Aug 1, 2026, 02:00 PM