Ascend Core

Trust-based agent orchestrator for building autonomous AI agent systems with safety guarantees.

What It Is

Ascend Core provides the execution layer for AI agent systems:

• Executor — Runs agent tasks through policy/trust/audit pipeline
• Trust Engine — SQLite-backed L0-L4 trust levels with auto-promotion/demotion
• Policy Engine — YAML-based action validation, blast radius control, forbidden paths
• Middleware — Budget tracking, loop detection, secret scrubbing, verification
• Memory — Entity store, exemplars, blackboard, trends, garbage collection
• Intelligence — Error classification, recovery strategies, prompt evolution

What It Is NOT

Ascend Core does not handle orchestration:

• No scheduling (use cron, Temporal, or your own scheduler)
• No messaging (use Telegram, Slack, or your own delivery)
• No webhooks (use your own HTTP server)
• No session management

These are handled by your orchestration layer. Ascend Core is the execution engine underneath.

Install

pip install ascend-core

Or from source:

git clone https://github.com/vitamin33/ascend-core.git
cd ascend-core
pip install -e ".[dev]"

Quick Start

from pathlib import Path
from ascend import TrustEngine, PolicyEngine, TaskContract
from ascend.executor import AgentExecutor

# Initialize engines
trust = TrustEngine(Path("data/trust.db"))
policy = PolicyEngine("config/policies.yaml", "config/safety.yaml")
executor = AgentExecutor(trust, policy)

# Create and execute a task
task = TaskContract(
    project="my-project",
    action="code_review",
    description="Review the latest PR for security issues",
    target_files=["src/auth.py"],
    trust_level=0,
)

result = executor.execute(task, agent_id="code-reviewer")
print(f"Status: {result.status}, Output: {result.output[:200]}")

See examples/minimal_daemon.py for a full HTTP server example.

Trust Levels

Level	Name	Can Do	Execution
L0	Observer	Read-only analysis	subprocess
L1	Contributor	Draft content, PR comments	subprocess
L2	Trusted	Auto-execute within policy	tmux
L3	Senior	Cross-project workflows	tmux
L4	Architect	System-level changes	tmux

Agents auto-promote after 10 consecutive successes and auto-demote after 2 failures within 24 hours.

Architecture

TaskContract
    |
    v
[Policy Engine] -- validates action, checks blast radius
    |
    v
[Trust Engine] -- checks agent trust level, decides execution mode
    |
    v
[Executor] -- runs agent via subprocess (L0-L1) or tmux (L2+)
    |
    v
[Audit Log] -- append-only JSONL, never deleted
    |
    v
AgentResult

Configuration

Two YAML files control behavior:

policies.yaml — What agents can do

projects:
  my-project:
    allowed_actions: [report, code_review, test]
    forbidden_paths: ["*.env", ".secrets/*"]
    blast_radius: low

agents:
  code-reviewer:
    allowed: [code_review, report]
    forbidden: [deploy, delete]

safety.yaml — Resource limits

timeouts:
  default_task_seconds: 300
  max_task_seconds: 900
budget:
  daily_api_calls: 500

See examples/ for complete configuration files.

Project Structure

ascend-core/
  ascend/
    __init__.py          # Package exports
    contracts.py         # TaskContract, AgentResult, WorkflowContract
    executor.py          # Core execution pipeline
    trust_engine.py      # SQLite trust L0-L4
    policy_engine.py     # YAML policy validation
    middleware/           # Budget, loop detection, secrets, verification
    memory/              # Entity store, exemplars, blackboard, trends, GC
    intelligence/        # Error classification, recovery, prompt evolution
  examples/
    minimal_daemon.py    # Working HTTP server (~90 lines)
    policy_example.yaml  # Sample policy config
    safety_example.yaml  # Sample safety config
  tests/
  docs/

License

Apache 2.0 — see LICENSE.

Author

Built by Vitalii Serbyn. Write-ups on agent reliability, with the numbers behind them: serbyn.io/writing.