Agent workforce overview

Aucert operates with an AI agent workforce where ~90% of task execution is autonomous. Agents are managed through Astra, the Agent Workforce Control Plane.

What is Astra?

Astra is the single source of truth for agent identity, credentials, configuration, and performance. Think of it as "the HR system for AI agents."

Key features (v0.1):

• Agent registry with full metadata and personality system
• Token vault (AES-256-GCM encrypted credential storage)
• Task run logger (append-only execution history)
• Console dashboard (workforce overview, agent detail, create wizard)
• Permission system (role-based, policy-gated)
• Audit log (every write operation tracked)

URL: astra.aucert.dev (Cloudflare Access, @aucert.ai only)

Active agents

Vega

Active

Vega · Coder · Thorough

Backend engineering — Kotlin, protobuf, database migrations

"Reads all context files before writing a single line"

Model: Claude Sonnet 4

Rigel

Active

Rigel · Coder · Fast

Frontend engineering — TypeScript, React, Next.js

"Ships working code quickly, iterates on review feedback"

Model: Claude Sonnet 4

Lyra

Active

Lyra · Docs · Creative

Documentation — architecture docs, how-to guides, ADRs

"Makes complex concepts accessible through diagrams and analogies"

Model: Claude Sonnet 4

Altair

Active

Altair · Ops · Thorough

Infrastructure — Terraform, Helm, CI/CD, AKS

"Triple-checks blast radius before any infra change"

Model: Claude Sonnet 4

Sirius

Standby

Sirius · Reviewer · Thorough

Code review — security, architecture, conventions

"Checks every PR against AI_RULES.md and CONVENTIONS.md"

Model: Claude Opus 4

Agent identity model

Each agent has:

Field	Description	Example
Fun name	Celestial-themed short name	Vega, Rigel, Lyra
Display name	Compound: {Fun Name} · {Role} · {Trait}	Vega · Coder · Thorough
Role	What the agent does	coder, ops, docs, reviewer
Trait	Behavioral modifier	thorough, fast, creative
Model	Which AI model powers the agent	claude-sonnet-4
System prompt prefix	Personality and behavioral guardrails	Context loading rules, phase boundaries

How agents execute tasks

1. Tasks created in Plane with capability labels (agent:coder, agent:ops)
2. Dispatcher matches task to available agent (based on role, capacity, skills)
3. Agent launches in isolated environment (K8s Job, planned for Phase 2)
4. Agent reads context files, code, and spec — then makes changes and raises PR
5. Results posted back to Plane
6. Human reviews and merges — agents never self-merge

Agent guardrails

All agents operate under the same rules as human developers:

• Read spec/ before modifying any module
• Never create Phase 2/3 directories
• Follow interface + adapter pattern
• Include context file updates in every infra/architecture PR
• Never store secrets in code
• Never merge their own PRs

What's next

• Dispatch architecture — How tasks reach agents
• Adding a new agent — Provisioning guide
• How to create an agent task — Task creation guide