Context Engineering · LLMOps · AI Token Controls

7 token guardrails.
Most teams ship 1–2.
Conduct enforces all 7.

A checklist is circulating in #ContextEngineering and #LLMOps. Your team probably covers two of these manually. Conduct enforces the full stack — across every developer and every AI tool — automatically from day one.

Set up Guard free → Book a demo
Works with Claude Code · Cursor · Windsurf · Codex 60–90% token reduction on common operations Per-developer spend visibility + hard caps Free during early access
The 7 guardrails

Every token that leaves your team
passes through this stack.

Each layer intercepts a category of waste. Most teams get one or two by accident. The rest run unchecked.

01
🗜️
Prompt caching
Repeated system prompts are sent in full on every call. Caching eliminates that cost on every agent run.
Conduct runtime cached
02
🔀
Model routing
Sending every task to Opus when Haiku would do costs 10–20× more per call. Conduct routes by task complexity at runtime.
Agent engine 10–20× cheaper
03
✂️
Mega-prompt splitting
Monolithic prompts bloat context windows and confuse models. Agent Templates enforce composable YAML skills — each step gets only what it needs.
YAML playbooks
04
⚙️
Deterministic offload
Tasks with known algorithms — sorting, formatting, regex — don't need a model. Guard's warn-deterministic-compute policy flags them before they reach the API.
Guard policy WARN
05
📦
Output compression
Git diffs, test results, and build logs are verbose by default. RTK strips noise before output is fed back as context — 60–90% reduction on common operations.
RTK 60–90%
06
🔍
Structured retrieval
Loading full files into context instead of relevant symbols is the single biggest waste in AI coding workflows. Agent Booster uses AST + vector retrieval to send only what's needed.
Agent Booster up to 15×
07
💸
Metrics & budgets
Without a hard cap, one runaway agent session can spike your monthly bill. Guard enforces per-developer limits — blocked at the source, not flagged after the fact.
Guard spend BLOCK
tokens reaching the model: minimal
Output compression (RTK)
60–90%
Token reduction on git, test, and build output before it re-enters context
Structured retrieval (Agent Booster)
up to 15×
Cost reduction per AI request by sending relevant symbols only, not full files
Guardrails covered automatically
7 / 7
From day one. Without rewriting your agents or changing your tools
Coverage map

What covers what.

Each guardrail maps to a specific Conduct component. No overlap, no gaps.

Guardrail Conduct component Enforcement Works with
Prompt caching Conduct runtime — system prompts cached on every agent run Automatic All Agent Templates
Model routing Agent engine selects model tier at runtime by task complexity Automatic All Agent Templates
Mega-prompt split YAML playbooks enforce composable skills — each step scoped Automatic All Agent Templates
Deterministic offload Guard warn-deterministic-compute policy — fires before model call WARN / BLOCK Claude Code · Cursor · Windsurf
Output compression RTK — strips noise from git, test, build, and file output Hook-level Claude Code · any CLI
Structured retrieval Agent Booster — AST parse + vector retrieval, relevant symbols only MCP tool Claude Code · Cursor · any MCP host
Metrics & budgets Guard spend dashboard — per-dev tracking, hard cap, booster gain --team Hard cap All tools via Guard
How to enable the full stack

Three installs. All 7 active.

Each component is independent. Install what you have time for today. Add the rest when ready. Guard shows which guardrails are active in real time.

🛡️

ConductGuard

Sets up spend limits, policy enforcement (deterministic offload, secrets, destructive ops), per-developer tracking, and the audit trail. Developers run conduct login — done.

conduct login # ✓ spend limit applied # ✓ policies enforced # ✓ audit trail active

Covers guardrails 4 and 7 · 5 mins to set up

📦

RTK — Rust Token Killer

Wraps common dev commands and compresses verbose output before it re-enters AI context. Drop-in: prefix any command with rtk. Or install the Claude Code hook for automatic coverage.

pip install rtk rtk git diff # 80% smaller rtk pytest # failures only rtk next build # 87% smaller

Covers guardrail 5 · 2 mins to install

🔍

Agent Booster

MCP server that indexes your codebase with AST parsing and vector search. AI tools call search_context and smart_read instead of reading full files — up to 15× cost reduction per request.

pip install agent-booster booster serve # starts MCP server booster index . # indexes repo

Covers guardrail 6 · 5 mins to install

◈

Conduct Agent Templates

18 pre-built YAML playbooks with prompt caching and model routing built in. Each step gets the right model and the right context — not a monolith prompt fired at Opus every time.

conduct install pr-reviewer # ✓ prompt caching on # ✓ model routing active # ✓ YAML skills composable

Covers guardrails 1, 2, 3 · 1-click install

📊

The Guard settings page shows which of the 7 guardrails are currently active for your workspace — auto-detected from tool installations and policies, updating in real time. No manual tracking needed.

"Guard paid for itself in week one. I could finally see exactly which developers and which tools were burning our Claude budget. We cut spend by 40% without slowing anyone down — and that was before we installed RTK."

VP Engineering
SaaS company · 60-person team

All 7. From day one.

Set up Guard in 5 minutes. Install RTK and Agent Booster while your team does stand-up. Run conduct login and every developer is covered.

Set up Guard free → Book a demo

For DevSecOps · LLMOps · VP Eng · No credit card required