Drafting Framework

The drafting framework implements multi-model generation strategies that trade off speed, quality, and cost. Strategies are defined in drafting/drafting_strategies.yaml.

Drafting is disabled by default (AgentConfig.enable_drafting = False).

Strategies

Speculative Decoding

A fast draft model generates tokens that a stronger target model verifies, accepting or rejecting each batch.

sequenceDiagram
    participant D as Draft Model (fast)
    participant T as Target Model (strong)

    loop Until done or max iterations
        D->>D: Generate N draft tokens
        D->>T: Send draft for verification
        T->>T: Score each token
        T-->>D: Accept/reject (threshold)
    end

Config	Description
`draft_model`	Fast model (e.g., `gpt-3.5-turbo`, `llama3.2`)
`target_model`	Strong model (e.g., `gpt-4-turbo`, `claude-3.5-sonnet`)
`draft_tokens`	Tokens per draft batch (20–30)
`acceptance_threshold`	Minimum score to accept (0.7–0.8)
`max_iterations`	Maximum draft-verify cycles

Pre-configured strategies: fast_accurate, local_cloud, claude_fast

Pipeline

Multi-stage generation where each stage uses a different model with a specific role.

Stage Role	Description
`analyze` / `code`	Initial generation
`critique` / `review`	Critical review
`refine`	Incorporate feedback
`summarize`	Final synthesis

Each stage has its own model, system prompt, and temperature.

Pre-configured strategies: code_review (generate → review → refine), writing_pipeline (outline → draft → edit → polish), analysis_pipeline (decompose → research → synthesize)

Candidate Generation

Generate multiple candidates and select the best using a scoring method.

Scoring Method	Description
`majority_vote`	Most common answer wins
`verifier`	Separate model scores each candidate
`length_preference`	Prefer longer/shorter responses

Pre-configured strategies: consensus (multi-model vote), best_of_n (N candidates + verifier), diverse_ensemble (varied models), self_consistency (same model, multiple samples)

Result Structure

DraftResult contains:

Field	Type	Description
`content`	string	Final output
`strategy`	string	Strategy name
`status`	DraftStatus	`"complete"` or `"failed"`
`draft_tokens`	int	Tokens drafted
`accepted_tokens`	int	Tokens accepted (speculative)
`models_used`	list[string]	All models involved
`stages_completed`	int	Pipeline stages run
`candidates_generated`	int	Candidates produced
`estimated_cost`	float	Estimated USD cost
`total_time_ms`	float	Elapsed time

Task Defaults

The defaults section in drafting_strategies.yaml maps task types to strategies:

Task	Strategy
`general`	`fast_accurate`
`code`	`code_review`
`writing`	`writing_pipeline`
`analysis`	`analysis_pipeline`
`consensus`	`consensus`

Providers — Model providers used by drafting
API Models: DraftResult — Result schema
Config file: api/agentx_ai/drafting/drafting_strategies.yaml