Behavioral Quality & UX

How the agent "feels" to interact with matters as much as what it can do. Autonomy calibration, model-specific behavioral quirks, response formatting, knowing when to be proactive vs. quiet, and having real opinions without being a contrarian — these are the patterns that make an agent feel like a competent collaborator rather than a verbose chatbot.

Key Problems

Effort-to-stakes mismatch

Quick question gets a 5-paragraph framework response. Complex request gets a one-liner. The agent struggles to match response depth to the actual stakes of the question. "What time is it?" doesn't need a methodology section.

Channel formatting is not one-size-fits-all

Telegram wants concise, no walls of text. Discord handles markdown but not tables well. Gists are for comprehensive, structured output. The agent needs different formatting instincts per channel.

Autonomy calibration

Too autonomous = takes actions without oversight, destroys trust. Not autonomous enough = asks permission for everything, useless. The sweet spot is a tiered system that expands through demonstrated competence.

Model selection affects behavior, not just capability

Different model families have distinct behavioral fingerprints — tool compliance, instruction following, writing quality, failure modes. Matching model to task based on behavioral characteristics matters as much as matching by raw capability.

Sycophancy

"Great question!" and "Absolutely!" before every response. Models default to agreement and praise. A useful agent has opinions, pushes back when the user is wrong, and doesn't pad responses with false enthusiasm.

Group chat behavior

In group contexts, the agent needs to know when to speak, when to react with an emoji, and when to stay completely silent. Over-participating is a fast way to become annoying.

What's Here

• Core Truths — The foundational beliefs that transform a request/response chatbot into an autonomous agent. Action bias, resourcefulness, opinions, trust-building, and the principles that guide behavior when no explicit rule applies.
• Multi-Tier Autonomy — Three-tier and four-tier autonomy models for single-agent and supervisory setups. How to define tier boundaries, earn trust over time, and avoid tier inflation/deflation.
• Model Behavioral Differences — How Claude, GPT, Gemini, and Codex variants behave differently in agent workloads. Tool compliance, writing quality, instruction adherence, and practical routing guidance.
• Streaming & Queue Mode — Config tuning for responsive Discord streaming and mid-run message interruption via steer mode.