System Architecture

Deprex operates on a daemon-driven autonomous architecture. Rather than waiting for humans to give instructions, the FDE continuously runs its own observe-learn-execute cycle.

The Autonomous Loop at the Center

Deprex’s architecture is designed around the following autonomous loop:

+--------------------------------------------------+
|            Autonomous Cognitive Loop             |
|                                                  |
|   Observe -> Learn -> Skillify -> Displace -> Improve |
|       ^                                     |    |
|       +-------------------------------------+    |
+--------------------------------------------------+
        |                    ^
        v                    |
+----------------+   +----------------+
| Input Layer    |   | Output Layer   |
| - Slack events |   | - Slack notify |
| - Cron trigger |   | - Files/reports|
| - Observation  |   | - PR/commits   |
+----------------+   +----------------+

Conventional AI tools follow a unidirectional “input -> process -> output” pattern. Deprex’s FDE has a circular cognitive loop that evaluates its own output and uses it for the next improvement.

7-Phase Cognitive Cycle

Internally, the FDE processes every task through a 7-phase cognitive cycle:

Phase	Role	Example
PERCEIVE	Observe current state	Understand Slack message content and channel context
REMEMBER	Recall relevant memory	Retrieve tenant KPI definitions and past report formats
REASON	Reason about the task	Determine which skills to use, identify missing information
PLAN	Formulate a plan	Compose execution steps, identify required resources
ACT	Execute	Run skills and generate results
EVALUATE	Evaluate results	Verify output quality, compare against ground truth
LEARN	Learn	Reflect feedback into skills and memory

This cycle does not run just once — it repeats recursively until quality converges.

Daemon-Driven Architecture

The FDE runs continuously via two daemon processes.

Slack Daemon

Receives Slack @mentions and begins real-time processing.

Incoming messages are automatically triaged into three levels:

• Reflexive response: Greetings or simple questions — answered immediately
• Light task: Summaries or lookups — processed and replied on the spot
• Complex task: Report generation or data analysis — fed into the cognitive loop

Triage is performed autonomously. PMs do not need to specify the task type.

Permanent Daemon

Executes scheduled tasks on cron schedules. Additionally, it performs autonomous maintenance:

• Memory Distill: Periodically organizes, consolidates, and detects staleness in accumulated memory
• Harness Autoresearch: Experiment-based optimization of skill parameters
• Specstory Distill: Extracts feedback patterns from session histories
• Autonomous Scheduling Optimization: Learns from each task’s execution history (success rate, average duration, failure patterns) and dynamically optimizes execution order. Scheduling precision improves with every run.

These run in the background without PM intervention.

Autonomous Task Displacement Pipeline

The FDE’s most important capability is autonomously displacing human tasks.

1. Workflow Observation
   The FDE detects task patterns from Slack conversations and meeting transcripts

2. Task Flow Recording
   Detected patterns are recorded as structured flows

3. Skill Auto-Generation
   SKILL.md + scripts are auto-generated from task flows

4. Shadow Comparison
   Human output and FDE output are run in parallel for quality comparison

5. Production Displacement
   After PM approval, the FDE switches tasks to automated execution

6. Continuous Improvement
   Skills are self-improved using execution feedback

This entire pipeline runs autonomously. PMs receive notifications at each stage and simply approve.

Multi-Tenant Isolation

To deploy FDEs to multiple client organizations simultaneously, tenants are completely isolated.

app/orgs/
  acme/          # Acme's dedicated environment
    skills/      # Skills auto-generated for Acme
    memory/      # Memory learned from Acme's workflows
    config/      # Acme-specific authentication and settings
  beta-inc/      # Beta Inc's dedicated environment
    skills/
    memory/
    config/

Each tenant’s memory, skills, and configuration are never accessed by other tenants. Business patterns observed in Acme’s Slack never leak into Beta Inc’s skills.

Self-Improvement Mechanisms

The FDE autonomously improves its own harness (skills, rules, hooks):

• Learning from failures: When task execution encounters problems, the FDE analyzes root causes and fixes skills
• Convergence with ground truth: Uses human output as ground truth, iterating improvements until quality is equal or better
• Feedback distillation: Extracts improvement patterns from session histories and reflects them in skills

All of this runs in the background without PM involvement. The only thing PMs notice is output quality gradually improving.

What This Means for PMs

PMs do not need to understand the architecture details. What matters:

1. The FDE is always running — observation and learning continue even without requests
2. Automation progresses incrementally — observe, propose, approve, execute — PMs maintain control
3. Quality improves autonomously — the FDE collects feedback and improves its own output continuously