4 key decisions to build reliable agent harnesses, from IBM engineer Tejas Kumar
TejasKumar_ · x · 2026-11-04
IBM AI engineer and Fluent React author Tejas Kumar breaks down the four ownership decisions in agent harness engineering: who decides a task is done, who controls sensitive actions, who says stop, and who's to blame when a run fails.
His core rule: the model proposes, the harness decides in code. Verify completion against an independent source of truth, keep credentials and irreversible operations outside the model's reach, separate budgets from blocking policies, and debug by tracing the first wrong step — bad context means fix the harness, correct context means measure the model.
He also revisits his 275k-view GPT-3.5 Turbo Hacker News upvoting demo: a community dev fixed a double-upvote bug, and rereading the code revealed the harness could be tricked into typing an HN password on a non-HN page — living proof of all four decisions.
More from coding & agent
- 8x Microsoft MVP ships complete Microsoft Agent Framework series and enterprise agents course — marlene_zw · 2026-11-04
- Living Brain tracks 3 health scores to stop AI agents relying on stale memory — cneuralnetwork · 2026-11-04
- Power user tallies 12B coding-agent tokens, finds 96% were cache reads — peakpirate007 · 2026-11-04
- Indian code-mixed calls have no single correct transcript: LLM, CRM and compliance each need a different one — BEYOND__7 · 2026-11-04
- Alexey Grigorev releases free kit of 28 prompts to ship an app from spec to production — Al_Grigor · 2026-11-04
- Hugging Face Hub MCP Server lets agents search models and datasets via MCP — modelcontextprotocol · 2026-11-04