Three layers. Zero loopholes. Kaidera controls worker behavior at structural, prompt, and activation levels simultaneously.
Every other platform puts a prompt in front of a model and calls it "steering." That's not control — that's hope.
The difference between "just tell it to be safe" and"enforce it architecturally" is the difference betweenhoping a worker behaves andknowing it must.
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 worker, every conversation, without exception. These constraints cannot be overridden by any individual worker configuration, template, or prompt. They are the non-negotiable policies of the organisation itself — enforced before the model ever sees a request.
Constraint Inheritance — Top to Bottom
Lower layers can only add restrictions — they can never remove a Floor-level lock.
Example Org Floor Policies
No worker touches production without an explicit human override
All write operations require human confirmation, org-wide
GDPR compliance flags enforced across every worker interaction
Workers restricted from calling external APIs without approval
The Constraint Layer
The deepest layer. Controls what a worker 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.
Worker State Machine — 7 States
AI Worker State Machine
Every state transition passes through the Impact Gate
Typos, tests, styling, comments
API, schema, shared components
Security, architecture, breaking
Impact Classification Gate
Every change is classified before execution proceeds
Examples
No approval needed. AI worker continues immediately.
Examples
Alert downstream workers. Log impact radius.
Examples
Hard stop. Chief Architect decides. Nothing proceeds.
The Kaidera Principle
Small radius = auto-process. Large radius = Chief Architect decides. Nothing breaks silently.
The Identity Layer
Every worker 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 workers always know exactly where they are and what they're allowed to do. Steering constraints are rendered asXML-structured templates — enabling the model 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.
Keith
CPO
Sophi
Backend
Marv
Frontend
PROMI
Orchestrator
Sage
Knowledge
Steering Preview Endpoint
Before deploying a worker, org admins callPOST /agents/{id}/steering/previewto see the fully-rendered system prompt — all layers composed, XML constraints included, and memory context injected exactly as the worker will receive it at execution time.
The Calibration Layer
The calibration layer. Creativity and compliance parameters — set per-worker and per-task. Design-mode workers like Keith and Sage get higher creativity for exploratory thinking. Execution workers 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
Trigger: fix, bug, typo, patch
Scope: 1–3 files
Skips planning. Immediate.
Trigger: add, simple, update
Scope: 1–10 files
Brief plan, then build.
Trigger: Default scope
Scope: 10–30 files
Standard dev loop.
Trigger: security, architecture
Scope: 30+ files
CTO approval required.
Layer 1 constrains. Layer 2 personalizes. Layer 3 calibrates. Together, they produce worker behavior that is simultaneously controlled, intelligent, and purposeful.
Layer Composition — Top Down
The deepest layer. Controls what AI workers can do — not just what they're asked to do.
The identity layer. YAML AI worker files rendered via Jinja2 — giving each AI worker their personality, knowledge, and constraints.
The calibration layer. Per-worker creativity and compliance tuning. Scale routing ensures the right workflow for the task size.
Controlled AI Worker Behavior
Constrained + Personalized + Calibrated
Before vs After 3-Layer Steering
Latest Release
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.
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
Kaidera is the only platform with true multi-layer behavioral steering. See how we stack up against other leading AI work platforms.
Platform Comparison
Behavioral steering capabilities across leading AI worker platforms
| Capability | Kaidera | 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 |
| AI Worker Identity Files | Full YAML | None | Basic | None | Basic |
Data based on publicly documented capabilities. Feb 2026.
Five structured phases. Verification gates at every step. How Kaidera workers prepare, research, execute, verify, and commit — systematically.
How PREVC Works
