Five concepts. That's all you need to understand GAAI.
GAAI separates two fundamentally different activities:
Discovery Track — figuring out what to build and why
- Human-facing. Produces artefacts. Asks questions. Validates intent.
- Owned by the Discovery Agent
Delivery Track — building it correctly and verifiably
- Execution-focused. Implements. Tests. Passes QA.
- Owned by the Delivery Agent
Discovery must complete before Delivery begins. This prevents the most common AI failure mode: building the wrong thing fast.
These two tracks are not just different phases — they are architecturally different systems:
| Dimension | Discovery Agent | Delivery Agent |
|---|---|---|
| Primary interlocutor | The human | The backlog |
| Nature of work | Conversation + clarification + artefact generation | Deterministic execution against a plan |
| Output | PRDs, Epics, Stories (~short, structured) | Code, tests, QA reports |
| Expected interruptions | Frequent — asks the human questions | Rare — only escalates when blocked |
| Context requirement | Conversational continuity is critical | Phase isolation improves quality |
| Authorization source | Human intent + project constraints | Backlog item with status: refined |
| When it stops | Never stops mid-conversation | Stops on QA failure or ambiguity |
This asymmetry is why Discovery and Delivery have different agents, different authority models, and different activation patterns. They are not interchangeable.
If it's not in the backlog, it must not be executed.
The backlog is not a to-do list. It is the only authority that permits the Delivery Agent to act.
- Location:
.gaai/project/contexts/backlog/active.backlog.yaml - Format: YAML (one file for all active items)
- States:
draft→needs-refinement→refined→in-progress→done
The Delivery Agent reads the backlog, picks the next refined item, and executes it. Nothing else.
Agents reason, decide, and orchestrate. They have authority.
Skills execute. They do one thing and produce explicit output. They never decide anything.
| Agent | Skill | |
|---|---|---|
| Interprets intent | Yes | No |
| Selects context | Yes | No |
| Decides strategy | Yes | No |
| Executes procedure | No | Yes |
| Produces artefacts | No | Yes |
The Delivery Agent invokes the implement skill. The implement skill cannot decide what to implement. That separation is what makes AI execution predictable.
If a skill appears to think, it is wrongly designed.
Memory in GAAI is never automatically loaded. Agents select what context they need. Always.
This matters because:
- Auto-loading all memory wastes tokens and introduces noise
- Context pollution leads to drift and contradictions
- Explicit selection makes agent behavior predictable and auditable
Memory lives in .gaai/project/contexts/memory/. The structure — project context, decisions log, patterns, summaries — is defined in .gaai/project/contexts/memory/README.memory.md. You can add domain-specific memory files; agents select them explicitly using memory-retrieve.
Artefacts (Epics, Stories, Plans) are documentation. They explain intent and decisions. They are not the authorization mechanism.
| Documents intent | Authorizes execution |
|---|---|
| Epic artefact | Backlog item |
| Story artefact | Backlog item with status: refined |
| PRD | Backlog items it generates |
An artefact with no backlog item cannot trigger execution. A backlog item with no artefact can still be executed (though artefacts are strongly recommended for clarity).
Human intent
↓
Discovery Agent (reasons, produces artefacts, validates)
↓
Backlog items (status: refined — authorization granted)
↓
Delivery Agent (invokes skills, executes plan, runs QA)
↓
Done items (move to done/)
↓
Memory captures decisions for future sessions
→ Getting Started — set it up → Quick Start — try it now