Technology Overview
Deep Dive — Behavioral Steering

The Architecture
of Trust

Three layers. Zero loopholes. The only platform that controls agent behavior at structural, prompt, and activation levels simultaneously.

Org Floor — Policy Lock
Layer 1 — Structural
Layer 2 — Prompt
Layer 3 — Behavioral

Why One Layer Isn't Enough

Every other platform puts a prompt in front of an LLM and calls it "steering." That's not control — that's hope.

The Prompt-Only Approach
"You are a helpful coding assistant. Be safe."
  • Agent ignores constraint when convenient
  • No domain boundary enforcement
  • Creativity vs safety is uncalibrated
  • Nothing stops hallucinated decisions
  • No human gate before shipping
The EnGenAI Approach
Structural constraints + Identity YAML + Calibrated parameters
  • 7-state lifecycle enforces what CAN happen
  • Impact gate halts HIGH-risk changes
  • Identity YAML personalizes behavior at render time
  • Creativity/compliance tuned per agent per task
  • Human approval gate before every merge

The difference between "just tell it to be safe" and "enforce it architecturally" is the difference between hoping an agent behaves and knowing it must.

0

Steering Floor

Org-Level Inviolable Policy

Below the three layers sits the Steering Floor — a set of locked keys defined by the org admin that apply uniformly to every agent, every conversation, without exception. These constraints cannot be overridden by any individual agent configuration, template, or prompt. They are the non-negotiable policies of the organisation itself — enforced before the LLM ever sees a request.

Constraint Inheritance — Top to Bottom

Steering Floor (Org Admin)
locked_keys — inviolable. Cannot be overridden by anything below
Layer 1 — Structural
State machine, impact gate, domain boundaries
Layer 2 — Prompt
Jinja2/XML templates, agent identity, memory injection
Layer 3 — Behavioral Parameters
creativity, compliance, execution_mode

Lower layers can only add restrictions — they can never remove a Floor-level lock.

Example Org Floor Policies

cannot_access_production = true

No agent touches production without an explicit human override

require_human_on_write = true

All write operations require human confirmation, org-wide

gdpr_compliant = true

GDPR compliance flags enforced across every agent interaction

domain_scope = "internal_only"

Agents restricted from calling external APIs without approval

1

Layer 1 — Structural

The Constraint Layer

The deepest layer. Controls what an agent can do — not just what it's asked to do. Seven lifecycle states. An impact classification gate that auto-proceeds, notifies, or escalates based on change radius. Hard domain boundaries that cannot be overridden by a clever prompt.

Agent State Machine — 7 States

Agent State Machine

IDLE Awaiting task
PLANNING Decompose task
EXECUTING Build & code
WAITING Pending approval
REVIEWING Quality check
COMPLETE Task shipped
3-strike rule
ERROR 3-strike halt
IDLE
PLANNING
EXECUTING
WAITING
REVIEWING
COMPLETE
ERROR
Impact Gate

Every state transition passes through the Impact Gate

LOW

Typos, tests, styling, comments

Auto-proceed
MEDIUM

API, schema, shared components

Notify areas
HIGH

Security, architecture, breaking

Escalate to CTO

Impact Classification Gate

Every change is classified before execution proceeds

LOW
RADIUS < 3

Examples

  • Typos & comments
  • Styling changes
  • Unit tests
  • Documentation
  • Log messages
Action
Auto-proceed

No approval needed. Agent continues immediately.

LOW
MEDIUM
RADIUS 3–10

Examples

  • API contract changes
  • Database schema updates
  • Shared components
  • Service interfaces
  • Config changes
Action
Notify affected

Alert downstream agents. Log impact radius.

MED
HIGH
RADIUS > 10

Examples

  • Breaking API changes
  • Security & auth logic
  • Architecture decisions
  • 3+ areas affected
  • Infrastructure changes
Action
Escalate to CTO

Hard stop. Chief Architect decides. Nothing proceeds.

HIGH

The EnGenAI Principle

Small radius = auto-process. Large radius = Chief Architect decides. Nothing breaks silently.

2

Layer 2 — Prompt

The Identity Layer

Every agent has a YAML identity file that gets rendered into their system prompt via Jinja2. This is how Keith knows he's CPO. How Sophi knows she can't touch the frontend. How Sage knows every decision needs documentation. Memory context, current phase, and active constraints are injected dynamically at every execution — so agents always know exactly where they are and what they're allowed to do. Steering constraints are rendered as XML-structured templates — enabling the LLM to reason over rules as structured data rather than freeform prose. This format reduces prompt token usage by 15–30% compared to Markdown-formatted constraints.

K

Keith

CPO

domain: product_strategy creativity: 0.8 compliance: 0.7
  • No direct backend code
  • Decompose before delegating
S

Sophi

Backend

domain: backend_services creativity: 0.4 compliance: 0.98
  • No frontend changes
  • Tests before code
M

Marv

Frontend

domain: frontend_ui creativity: 0.75 compliance: 0.85
  • No DB schema changes
  • WCAG 2.1 AA minimum
P

PROMI

Orchestrator

domain: orchestration creativity: 0.5 compliance: 0.95
  • Monitor all agents
  • Never bypass Impact Gate
S

Sage

Knowledge

domain: documentation creativity: 0.85 compliance: 0.75
  • 1 file per subject
  • Update on every change
Identity YAML
Jinja2 Engine
Memory Injection
System Prompt

Steering Preview Endpoint

Before deploying an agent, org admins call POST /agents/{id}/steering/preview to see the fully-rendered system prompt — all layers composed, XML constraints included, and memory context injected exactly as the agent will receive it at execution time.

3

Layer 3 — Activation

The Calibration Layer

The calibration layer. Creativity and compliance parameters — set per-agent and per-task. Design-mode agents like Keith and Sage get higher creativity for exploratory thinking. Execution agents like Sophi and Marv run near-maximum compliance to eliminate deviation. Scale routing automatically selects the right workflow depth: QUICK tasks skip planning, LARGE tasks require CTO approval and full sprint-closing.

Scale-Adaptive Task Routing

Q
QUICK

Trigger: fix, bug, typo, patch

Scope: 1–3 files

Execute
Verify

Skips planning. Immediate.

S
SMALL

Trigger: add, simple, update

Scope: 1–10 files

Plan
Execute
Verify

Brief plan, then build.

M
MEDIUM

Trigger: Default scope

Scope: 10–30 files

Plan
Review
Execute
Verify

Standard dev loop.

L
LARGE

Trigger: security, architecture

Scope: 30+ files

Plan
Approve
Execute
Verify
Confirm

CTO approval required.

How the Three Layers Compose

Layer 1 constrains. Layer 2 personalizes. Layer 3 calibrates. Together, they produce agent behavior that is simultaneously controlled, intelligent, and purposeful.

Layer Composition — Top Down

1
STRUCTURAL CONSTRAINS

The deepest layer. Controls what agents can do — not just what they're asked to do.

7-State LifecycleImpact GateTool GatingDomain ScopeTeam Hierarchy
constrains ↓
2
PROMPT PERSONALIZES

The identity layer. YAML agent files rendered via Jinja2 — giving each agent their personality, knowledge, and constraints.

Agent Identity YAMLJinja2 EngineMemory InjectionPhase AwarenessSystem Prompt
personalizes ↓
3
ACTIVATION CALIBRATES

The calibration layer. Per-agent creativity and compliance tuning. Scale routing ensures the right workflow for the task size.

Creativity ParameterCompliance ParameterExecution ModeScale RoutingCircuit Breaker

Controlled Agent Behavior

Constrained + Personalized + Calibrated

Before vs After 3-Layer Steering

Without 3-Layer Steering
"Build a login page"
  • Hallucinates framework choice
  • Uses wrong tech stack
  • Skips test coverage
  • No branch — commits to main
  • No PR — no human review
  • No security checks
With 3-Layer Steering
"Build a login page"
  • Checks domain scope first
  • Verifies stack from identity YAML
  • Follows PREVC — Prepare → Execute
  • Creates feature branch automatically
  • Writes tests before code (TDD)
  • Opens PR, waits for human merge

Latest Release

Steering Efficiency Dashboard

Org admins can measure the ROI of XML-structured steering in real time. The analytics endpoint returns three metric groups: XML vs Markdown token savings percentage, per-provider routing cost and latency, and observational memory compression stats. Set a lookback window from 1 to 365 days.

xml_token_savings_pct
per-provider cost + latency
observation compression stats
SAMPLE METRIC
21.5%
XML token savings
vs Markdown format

Interactive Explorer

Click through each layer to see how they work together. Explore the lifecycle states, identity files, and calibration parameters.

Click any lifecycle state to see what happens inside it.

Select a state above to expand its detail

Impact Gate

LOW
  • Typos & comments
  • Styling changes
  • Unit tests
  • Documentation
Auto-proceed
MEDIUM
  • API changes
  • Schema updates
  • Shared components
  • Config changes
Notify affected areas
HIGH
  • Breaking changes
  • Security & auth
  • Architecture decisions
  • 3+ areas affected
Escalate to CTO

How We Compare

EnGenAI is the only platform with true multi-layer behavioral steering. See how we stack up against other leading agent frameworks.

Platform Comparison

Behavioral steering capabilities across leading agent platforms

Capability
EnGenAI
Claude Code CrewAI LangGraph AutoGen
Steering Layers 3 layers 1 (prompt) 1 (prompt) 1 (prompt) 1 (prompt)
Memory Tiers 3 tiers 0 1 1 1
Human Approval Gate Always Never Optional Optional Optional
Real-time Canvas
Skill Vetting Pipeline 5-stage N/A None None None
Agent Identity Files Full YAML None Basic None Basic
EnGenAI advantage
Not supported

Data based on publicly documented capabilities. Feb 2026.

Next Deep Dive

Context Engineering

Five structured phases. Verification gates at every step. How EnGenAI agents prepare, research, execute, verify, and commit — systematically.

How PREVC Works

See It In Action

We're onboarding select B2B teams for early access. Tell us about your company and we'll be in touch.

Register for Early Access →