MOC

A map of content for the knowledge wiki. Each entry is an atomic, claim-titled note. Browse at /wiki/<slug>.

The nature of agents

• Agents amplify expertise rather than replace it
• Comprehension debt compounds silently
• Friction in agentic coding is engineering judgment
• Tenacity is a core bottleneck agents lift
• Supervised agents outperform fully autonomous ones

How context works

• Context management is the dominant variable in agent quality and cost
• Repo maps let agents navigate large codebases
• Deep modules make a codebase explorable and testable for agents
• Sub-agents keep the primary context focused
• Specs become the bottleneck once code is cheap
• Shared organizational context is the real limit

How to work with agents

• Spec before code for better agent output
• Break work into small iterative chunks
• Workflows encode domain knowledge better than agents
• Agent experience is developer experience for agents
• BYOK gives control over cost and privacy
• Commit often and never ship code you cannot explain
• Tests and CI are force multipliers for agent quality
• The rate of feedback is an agent’s speed limit
• A local cloud emulator tightens the agent feedback loop before deployment
• Persistent memory lets corrections compound

How agents connect

• MCP is a protocol layer that decouples tools from providers
• Code intelligence replaces dozens of grep calls with one graph query
• Multi-agent review beats single-prompt review
• Self-improving review systems adapt from developer feedback

How agents are trusted

• Tracing and evaluation serve different jobs
• Evals define what success means for an agent
• A hard-to-eval product is hard for users to verify
• Structure in structure out for LLM calls
• pass at k and pass to the kth measure opposite qualities

The knowledge layer

• An LLM wiki compounds where RAG rediscovers
• Git-native workflows make every change a checkpoint

Security

• the lethal trifecta makes agents exploitable
• guardrails alone cannot prevent prompt injection
• memory can become a vector for persistent attacks

Model selection and cost

• model choice trades quality for cost and speed
• token economics dominate the cost of agent projects

Production deployment

• agent actions and planning must be observable
• agents need human confirmation for consequential actions

Regulated environments

• regulated environments demand deterministic evidence for non-deterministic systems
• eval suites are qualification protocols in regulated settings
• human-in-the-loop is the compliance bridge for AI in pharma
• distribution shift is the silent regulator in production AI

Agent memory

• agent memory persists context that the model window cannot hold
• file-based memory compounds corrections across sessions
• memory is part of the harness not a separate tool
• OKF gives every agent durable company knowledge from a folder of typed markdown
• an LLM gateway sits between your app and providers

How agents are defined

• an agent harness is everything that constrains and shapes an agent
• agent skills are reusable named procedures an agent loads on demand
• a terminal multiplexer keeps many agents alive at once

Building with agents

• one working reference example beats documentation for agent-assisted builds
• expose app functionality as tools the agent can call
• inference is profitable training is what costs

Unsorted

• Eval systems create a flywheel for iteration
• Remove all friction from looking at data