Agent harnesses: why human-in-the-loop is the 2026 architecture, not a compliance checkbox

Loading map…

Agent harnesses: why human-in-the-loop is the 2026 architecture, not a compliance checkbox — Context map

#15 · EpicAgent harnesses: why human-in-the-loop is the 2026 architecture, not a compliance checkboxDone

Goal

What success means

How this was shaped

Interview transcript

Aug 29

Constraints & scope

Out of scope

Decisions

Decision record

Aug 29

Issue	Latest decision	Who / when
#17 Plan review	Approved	Human reviewer · 8/30/2026, 1:52:10 AM
#28 Package: PDF + HTML preview	Approved	Human reviewer · 8/30/2026, 2:19:19 AM

How we will execute

Delivery is complete

6 of 6 leaf issues complete

Agents, skills & reusable setup

1 model · 1 skill · tools: sourcing sweep — Claude Code, cheap model, low reasoning, target ≤$0.5… · gates: plan, final

Reported usage covers 5 of 6 issues. Unreported effort and cost are unknown.

6/6 done · 0 active · 0 blocked · 0 backlogagent— · — tokens · — · claude-sonnet-5human154 min · 2 decisionscreatedAug 29

Explore agents, models, skills and review gates

Driver & models

No agent reported `role: "driver"` — the rows below are who did the work.

Harness	Model	Thinking	Issues	Tokens	$
unknown	claude-sonnet-5	low	3	—	—
unknown	claude-sonnet-5	high	2	—	—

Skills

research

Aug 29

Tools & models guidance

sourcing sweep — Claude Code, cheap model, low reasoning, target ≤$0.50; outline + full draft — strongest model available, extended thinking, target ≤$2; fact-check pass — OpenCode on Kimi K2 against sources.md, target ≤$0.30; PDF/HTML packaging — Claude Code, python/markdown→PDF (no extra model cost); hard total token-cost cap $3.