[observability escalation] Design Decision Gate 🏗️ — repeated resource-heavy runs with increasing turn counts

[github-actions]

Design Decision Gate 🏗️ has crossed the escalation threshold with 2 consecutive runs both carrying a high-severity resource_heavy_for_domain assessment, cohort-matched baselines confirming turns_increase, and the latest run also showing medium poor_agentic_control.

Evidence

Run	Turns	Classification	Assessments
§24648171195	4	baseline	—
§24649877750	8	changed / turns_increase	resource_heavy_for_domain HIGH
§24650414222	10	changed / turns_increase	resource_heavy_for_domain HIGH, poor_agentic_control MEDIUM

The turn count has grown 4 → 8 → 10 across 3 runs today. The latest run used 9 distinct tool types in an exploratory, broad-tool posture. Agentic fraction is ~50%, meaning roughly half the turns are pure data-gathering that could be moved to deterministic pre-steps.

Why This Matters

• Engine: claude (frontier model cost)
• Execution style: exploratory, tool breadth: broad — weakly scoped
• The system flagged it as partially_reducible and model_downgrade_available
• Pattern is worsening across runs, not stabilizing

Recommended Fix

1. Tighten instructions — Reduce scope so the agent does not explore broadly. Narrow the task to a specific question rather than open-ended analysis.
2. Move data-gathering to pre-steps — The ~50% data-gathering fraction can become deterministic frontmatter steps writing to /tmp/gh-aw/agent/, reducing inference turns.
3. Reduce tool breadth — Limit available tools to what the task strictly needs. 9 tool types for an issue-response workflow is excessive.
4. Try a smaller model — claude-haiku-4-5 is recommended by the assessment. The task (issue response / decision gating) is well-suited to a smaller model.

Suggested Route

workflow:Design Decision Gate 🏗️ — workflow owners should review the prompt and tool configuration.

See the full observability report discussion for context on related workflows and the broader resource-heaviness pattern across the repository.

Generated by Agentic Observability Kit · ● 887.4K · ◷