Agent Orchestration Pipeline

The ASDD pipeline is a deterministic, confidence-gated sequence of 10 specialized agents. Each agent has a defined input, a defined output, a confidence threshold, and a failure protocol. No agent operates outside these boundaries.

---

The 10 agents at a glance

#	Agent	Role	Min Confidence
0	Discovery Agent	Converts intent into behavioral capabilities	0.85
1	Spec Agent	Formats capabilities as EARS requirements	0.85
2	Validation Agent	Detects ambiguity, validates domain terms, scores testability	0.90
3	Domain Agent	Builds and validates the domain model	0.85
4	Design Agent	Synthesizes architecture from requirements + domain	0.80
5	Task Planning Agent	Decomposes design into execution waves	0.80
6	Implementation Agent	Orchestrates code generation via context-fresh sub-agents	0.75
7	QA Agent	Validates spec coverage and test coverage	0.85
8	Refactor Agent	Maintains code quality within spec boundaries	0.80
9	Knowledge Agent	System memory, state, conflict detection, learning proposals	0.80

Plus:

• Security Agent — Pre-deployment compliance scan (min confidence: 0.95)
• DevOps Agent — CI/CD gate automation
• Observability Agent — Telemetry validation

---

The Cumulative Confidence Score (CCS)

The CCS is the most important quality mechanism in ASDD.

How it's calculated

CCS = Conf_Discovery × Conf_Spec × Conf_Validation × Conf_Design × Conf_Implementation

Example:

• Discovery Agent: 0.90
• Spec Agent: 0.88
• Validation Agent: 0.91
• Design Agent: 0.82
• Implementation Agent: 0.78

CCS = 0.90 × 0.88 × 0.91 × 0.82 × 0.78 = 0.455 → Pipeline halts, full human review required.

Each agent may individually pass its minimum threshold, but their combined uncertainty can still be unacceptable. The CCS catches this "cascading confidence failure" that individual gates miss.

Dynamic gating

When a preceding agent provides a low (but passing) score, the next agent's minimum threshold is increased by +0.05:

If Spec Agent scores 0.86 (above 0.85 minimum, but close to the floor):
  Validation Agent minimum increases from 0.90 → 0.95 for this slice.

This compensates for accumulated "uncertainty debt" in the pipeline.

Uncertainty factors

Any agent scoring below 0.95 must emit an uncertainty_factors list alongside its artifact. This allows the Tech Lead to identify the weak link in a low-CCS slice without reading every artifact in detail.

Example:

confidence_score: 0.78
uncertainty_factors:
  - "The 'Settlement' entity has conflicting definitions in two bounded contexts"
  - "The NFR for p95 latency is not specified in the requirements"
  - "The authentication flow references a token type not in the domain model"

---

Agent failure protocol

Agent failure is a guaranteed operational condition, not an edge case.

Confidence threshold action table

Agent	Min Confidence	Action if Below
Discovery Agent	0.85	Status → DRAFT — AWAITING CLARIFICATION; route to PO/TL
Spec Agent	0.85	Flag ambiguous sections; route to TL
Validation Agent	0.90	Block pipeline; require TL sign-off
Domain Agent	0.85	Draft domain model; require TL review
Design Agent	0.80	Draft architecture; require TL review before proceeding
Implementation Agent	0.75	Commit to feature branch; require human code review
Security Agent	0.95	Block deployment; require TL + Security review
Knowledge Agent	0.80	Propose steering update; require human approval

Failure modes and responses

Failure Mode	Detection	Response
Hallucinated spec section	Validation Agent detects undefined domain terms	Block pipeline; flag section; notify TL
Conflicting agent outputs	Design Agent output contradicts domain contract	Pause pipeline; log conflict; escalate to TL
Agent timeout / crash	Pipeline monitor: no output after N minutes	Retry once; then escalate to TL
Contradictory steering rules	Security Agent detects rule conflict	Block deployment; log conflict; notify TL
Infinite refinement loop	Same artifact version 3+ times	Force-halt pipeline; escalate to TL
Cascading Confidence Failure	CCS < 0.65	Halt pipeline; full human review of all artifacts in slice
Low-confidence cascade	Two or more consecutive agents below threshold	Halt pipeline; full human review of phase

Rollback procedure

When a pipeline halt is triggered:

1. The current phase's artifact is marked DRAFT — AGENT HALTED
2. All downstream agents are paused
3. The Tech Lead receives an automated escalation notice with: confidence score, conflicting artifacts, failure mode category
4. TL either resolves manually, returns to earlier phase, or approves a revised agent run
5. Resolution decision and rationale logged in /docs/agent-failure-log.md

Agent Failure Log schema

Date | Agent | Phase | Failure Mode | Confidence Score | Resolution | Time to Resolve | Root Cause

This log feeds the Knowledge Agent's learning loop.

---

Context-fresh sub-agents

The Implementation Agent spawns context-fresh sub-agents — isolated agents with clean context windows — for each implementation task. This prevents context rot.

Why context matters

Agents degrade in long, accumulating context windows. An Implementation Agent that has been running for 20 tasks accumulates domain knowledge, error states, and intermediate decisions in its context window. This degrades the quality of subsequent outputs.

Context-fresh sub-agents solve this: each task gets a minimal, purposeful context containing only:

• The specific task definition
• The relevant spec requirement (REQ-NNN)
• The local files relevant to the task
• The applicable domain terms from ubiquitous_language
• The relevant steering rules

Everything else is withheld.

The orchestrator-executor pattern

Implementation Agent (Orchestrator)
  └── Wave 1 tasks → Spawns Fresh Sub-Agent A (migrations)
  └── Wave 1 tasks → Spawns Fresh Sub-Agent B (domain models)
  └── Wave 2 tasks → Spawns Fresh Sub-Agent C (repositories)
  └── Wave 2 tasks → Spawns Fresh Sub-Agent D (domain services)

Sub-agents run in parallel within each wave. The Orchestrator coordinates sequencing between waves (Wave 2 starts only after Wave 1 is complete).

Goal-backward verification

Sub-agents are explicitly instructed to verify their work against the original business requirement (REQ-NNN), not just the technical task. This prevents "technically correct but spec-noncompliant" outputs.

---

Assumptions-First Discovery Mode

The Discovery Agent can operate in Assumptions-First Mode to accelerate Phase 0:

1. Agent analyzes intent.md and generates Initial Assumptions about system behavior and architecture
2. Presents to PO/TL: "I assume we are using X library and Y database structure. Correct me if I'm wrong."
3. Human only intervenes when an assumption is incorrect
4. Pipeline accelerates directly to requirements.md generation

This dramatically reduces "chat latency" — the back-and-forth questioning cycle that slows traditional discovery.

---

Agent-Agent conflict resolution

When parallel agents produce conflicting outputs (overlapping domain model modifications, contradictory architecture proposals), the resolution sequence is:

1. Knowledge Agent detects conflict via State Manifest
2. Knowledge Agent locks the affected slice (status → BLOCKED)
3. Knowledge Agent attempts resolution using Steering Rules
4. If resolved: Knowledge Agent unlocks slice; pipeline resumes
5. If unresolved within threshold: Tech Lead notified with conflict diff
6. Tech Lead arbitrates; decision logged in manifest

Unresolved agent-agent conflicts that do not escalate to a human are governance failures, not technical ones.

---

The ASDD Quick Mode (/asdd:quick)

For trivial tasks, the full pipeline is bypassed while maintaining traceability:

/asdd:quick "Fix typo in login header"

1. Intent parsed from one-line description
2. Implementation Agent generates fix directly (no spec phase)
3. QA Agent verifies against codebase
4. Knowledge Agent logs as QUICK_FIX slice in manifest.json

Scope restriction: Quick Mode applies only to tasks with zero architectural impact. The Tech Lead defines the boundary.

---

Next

• Governance — the full confidence scoring, dissent protocol, and CCS guardrail specifications
• Agent Catalog — individual agent specifications
• State Management — how the State Manifest tracks all agent activity