Evidence-first growth systems

Diagnose what moves growth. Fix the right lever.

Essays, formulas, and no-login tools for deciding what to change first. TokenMaxer starts with one likely bottleneck, then gives you one reading path and one workbench.

Friction → Bad CPA → CPA Workbench

If traffic is not the real problem, stop treating CPA as a single score. Map the constraint before changing spend.

Latest essays

16 published
Confidence AI Agent Control Gates: Stop Bad Agents Before They Act A practical control layer for production agents: monitoring, observability, tracing, security, MCP authentication, governance, evaluation, testing, human approval, and prompt caching. Jun 17, 2026 · 8 min Confidence Agent Observability: Trace What Agents Decide and Do A production agent observability guide with trace schemas, failure modes, and monitoring/security handoffs. Jun 17, 2026 · 8 min Confidence Agent Testing: Test Plans for Tool-Using AI Systems A practical agent testing plan for tool-using AI systems with scenarios for prompts, tools, retrieval, permissions, approvals, and budget stops. Jun 17, 2026 · 8 min Confidence Agent Tracing: A Practical Schema for Tool-Using AI A practical agent tracing schema for teams building tool-using AI agents, with fields for turns, tools, policy, cost, evals, and approvals. Jun 17, 2026 · 8 min Confidence AI Agent Evaluation: Gates That Catch Bad Behavior A practical AI agent evaluation guide with an eval gate for task success, tool use, groundedness, policy compliance, and regression testing. Jun 17, 2026 · 8 min Confidence AI Agent Governance: Approval, Audit, and Eval Gates A practical governance framework for production AI agents, with owners, risk tiers, approval paths, eval gates, and audit requirements. Jun 17, 2026 · 8 min Confidence AI Agent Monitoring: Metrics, Logs, and Stop Conditions A practical monitoring checklist for production AI agents: runtime metrics, loop detection, tool-call errors, budget stops, cache regressions, and escalation rules. Jun 17, 2026 · 8 min Confidence AI Agent Security: Threat Models for Tool-Using Agents A security threat model and checklist for production AI agents that call tools, touch private data, and trigger external actions. Jun 17, 2026 · 8 min Confidence Human Approval for AI Agents: When Agents Should Stop A human-approval matrix for AI agents that need to send messages, move money, deploy code, change data, or take customer-visible actions. Jun 17, 2026 · 8 min Confidence LLM Observability: When Basic Telemetry Stops Working An LLM observability guide that explains where model telemetry ends and agent observability begins. Jun 17, 2026 · 8 min Confidence MCP Authentication: Gate Agent Access to Tools Safely An implementation-oriented MCP authentication guide for teams connecting agents to restricted servers and sensitive tools. Jun 17, 2026 · 8 min Confidence MCP Security: Review Checklist for Agent Tool Servers A practical MCP security checklist for teams connecting AI agents to tool servers, private resources, and external actions. Jun 17, 2026 · 8 min Confidence Prompt Caching: Cut Agent Cost Without Breaking Quality A production-agent prompt caching guide for reducing cost without breaking context, eval, and trace behavior. Jun 17, 2026 · 8 min Friction Bad CPA Is Not a Diagnosis: A Constraint Map for Meta Ads A practical Meta ads operating model: decompose CPA into price, friction, value, and confidence, then use bubble charts to decide what to scale, fix, lower, or kill. Jun 16, 2026 · 18 min Confidence Better Embeddings for Vector Search Start Before the Database A practical argument for better embeddings: use LLMs to compile raw evidence into semantic cards, query probes, hard negatives, trust metadata, and eval harnesses. Jun 16, 2026 · 13 min Confidence Agent Harness Architecture: 15 Jobs Around the Model Loop A practical architecture guide for production agents: decompose the harness into replaceable jobs, design prompt caching into context layout, enforce tool policy in code, separate sessions from world state, and gate irreversible work. Jun 16, 2026 · 16 min

Tools attached to the reading

All tools

Four constraints, shown as rows.