#Agents

Engineering Decision Skills: What I Want AI to Do in Reviews, Planning, and Retrospectives

A proposal is due for review, three projects do not fit the next planning cycle, and another delivery is late. Five independent skills, each with an input and a concrete output you can take into a discussion.

RelationForge: Make Sense of an Interaction Before You Respond

Some relationships leave you uneasy, but you cannot quite say why. RelationForge can help you make sense of what happened, including your part in it.

Beyond Model Switching: Engineering Design and Routing Algorithms in DSH Quota Router

How candidate chains, failure classification, per-turn identity, and cost formulas make multi-source model routing controllable and auditable.

Let the Agent Execute: Emotional Adaptation, Trust Calibration, and Human Relief from Execution Pressure

Based on long-task use, this paper examines how an Agent can adapt state, evidence, and action feedback to uncertainty, waiting, failure, and takeover needs, while documenting how delegated execution changes human pressure, trust, and capability boundaries.

13 min read

Making an Agent a Collaborable Object: State, Feedback, and Result Confirmation

Based on ChatLab design work, this paper records how presence, response, waiting, and delivery patterns from human collaboration can make Agent runtime states easier to understand; it is an HCI design observation about feedback, evidence, intervention, and result confirmation.

13 min read

Quota Router: An Explainable Multi-Source Fallback Chain for DSH

A policy-only DeepSeek Harness plugin that builds ordered, explainable, auditable model candidate chains per task and advances along them within bounded failure rules.

Coding Agent Harness Study (I): Architecture Comparison and Restricted File Execution

Starting from public material, observable behavior, and a runnable microkernel, learn how to turn a coding agent's permissions, tool calls, and execution traces into a verifiable system boundary.

CanonLoom

A command-driven workflow for long-form novel production: the author drives with short commands, an Agent does the creative work, Python performs deterministic validation, and the author makes choices and approvals at every key checkpoint.