[agentic-token-optimizer] Daily Agentic Workflow AIC Usage Audit — Prompt & Chart-Phase Optimization

[github-actions]

🎯 Target Workflow

Daily Agentic Workflow AIC Usage Audit (agentic-token-audit)
Selected as the highest-AIC workflow not optimized in the last 14 days (last entry: 2026-07-27, 8 days ago). All other top-5 workflows were optimized within the past 7 days.

📅 Analysis Period

2026-07-29 → 2026-08-04 · 5 runs analyzed

📊 Spend Profile

Metric	Value
Runs analyzed	5
Total AIC	924.16
Avg AIC / run	184.83
Total tokens	574,378 (1-run sample)
Avg turns / run	3.0
Conclusions	5 × success (100%)
Cache efficiency	N/A (token breakdown not available for 4/5 runs)

Avg 184.83 AIC/run is ~2× the "Daily Agentic AI Research Digest" (94 AIC/run) despite this workflow having only 3 avg turns and a more structured, data-processing task. The primary cost driver is prompt size — the workflow prompt is 333 lines and carries significant in-prompt reference documentation that the agent re-reads every turn.

References: §30911974481 · §30817584209 · §30632873627

---

🔧 Ranked Recommendations

1. Remove the full RunData field reference table from the prompt

Estimated savings: ~20–30 AIC/run

The ## Data Sources → Pre-downloaded logs section embeds a 12-row Markdown table documenting every field of the RunData object (workflow_name, workflow_path, aic, token_usage, effective_tokens, action_minutes, turns, duration, created_at, run_id, url, status, conclusion, error_count, warning_count, token_usage_summary). This table is pure reference material — the agent only ever uses 6–7 of these fields (workflow_name, aic, token_usage, turns, conclusion, action_minutes, error_count).

Action: Replace the full 12-field table with a compact note listing only the fields actually used in the Python script:

Key fields per run: `workflow_name`, `aic` (float, treat null as 0), `token_usage` (int, treat null as 0),
`turns`, `conclusion`, `action_minutes`, `error_count`, `warning_count`, `url`. All other fields can be ignored.

This removes ~20 lines of prompt text that is re-loaded every turn.

---

2. Extract the repeated PYTHONPATH instruction into a setup note

Estimated savings: ~5–8 AIC/run

The ## Phase 3 section mentions the PYTHONPATH prefix twice — once in the bullet list and once in the inline code example. The same instruction was also present in the chart-generation context from earlier runs. Consolidating this into a single ## Setup Note near the top of the prompt ensures the agent picks it up once rather than re-encountering it mid-task.

Action: Add a short setup block after ## Data Sources:

## Setup Note
All Python commands that import pandas, matplotlib, or seaborn must be prefixed with:
`PYTHONPATH=/tmp/gh-aw/token-audit/site-packages${PYTHONPATH:+:$PYTHONPATH}`

Then remove the repeated PYTHONPATH callout from Phase 3.

---

3. Trim Phase 3 chart requirements

Estimated savings: ~10–15 AIC/run

The chart generation section (## Phase 3 — Generate Charts) is ~60 lines long and includes:

• Detailed styling requirements (300 DPI, white background, whitegrid, axis labels)
• URL replacement instructions for two named placeholders
• A conditional skip rule for when rolling-summary has fewer than 2 points
• Separate bullet items for each chart

Most of this detail could be condensed. The agent succeeds in 100% of runs, meaning it does not need this level of scaffolding.

Action: Reduce Phase 3 to the essential contract:

## Phase 3 — Generate Charts

Write and run Python scripts to generate PNG charts (300 DPI, white background, seaborn whitegrid):

1. **`ai_credits_by_workflow.png`**: horizontal bar chart, top 15 workflows by total AI credits from `audit_snapshot.json`.
2. **`ai_credits_trend.png`**: dual-axis line chart from `rolling-summary.json` — total AI credits (left axis) and `active_workflows` per day (right axis, label: "Active workflows/day"). Skip if fewer than 2 data points.

Upload each chart with `upload_asset`. Use the returned URLs as inline images in the issue.

This cuts ~35 lines while preserving all required behavior.

---

4. Collapse the issue template boilerplate

Estimated savings: ~5–8 AIC/run

## Phase 4 — Publish Audit Issue embeds a verbatim Markdown template (~50 lines) including the ### 📊 Executive Summary heading structure. The template uses named placeholders (UPLOAD_URL_WORKFLOW_PLACEHOLDER, etc.) that the agent must track across turns. A more concise instruction with a short schema description would be sufficient.

Action: Replace the full template with a structured description:

Create an issue with:
- Executive summary: period, total runs, total AIC, total tokens, total action-minutes, active workflows
- Top 5 table: workflow name, runs, total AIC, avg AIC
- Inline chart images from upload_asset URLs
- Per-workflow detail table (collapsible)
- Final observations (2–4 bullets on trend, top spenders, anomalies)

---

🔬 Structural Optimization: Inline Sub-Agent for Chart Generation

Phase 3 (Generate Charts) — sub-agent candidate

The chart generation phase is fully file-I/O driven: it reads two JSON files (audit_snapshot.json, rolling-summary.json), writes two PNG files, and uploads them. It requires no strategic reasoning, only procedural Python execution.

Dimension	Score	Rationale
Independence	2/3	Depends on Phase 1+2 output files (available before Phase 3 runs)
Small-model adequacy	3/3	Purely procedural: write Python → run → upload
Parallelism	1/2	Sequential (after Phase 2), but isolated
Size	2/2	Substantial enough to justify agent call
Total	8/10	Strong candidate

Proposed change: Replace the ## Phase 3 instructions in the main prompt with:

## Phase 3 — Generate Charts

## agent: chart-generator
Generate two PNG charts from pre-computed audit data and upload them. Return the upload URLs.

Input files (already written to disk):
- `/tmp/gh-aw/token-audit/audit_snapshot.json` — per-workflow AIC totals
- `/tmp/gh-aw/repo-memory/default/rolling-summary.json` — historical daily totals

Tasks:
1. Write and run a Python script (PYTHONPATH=/tmp/gh-aw/token-audit/site-packages) using pandas + matplotlib + seaborn (whitegrid, 300 DPI, white background):
   - `ai_credits_by_workflow.png`: horizontal bar chart, top 15 workflows by total_ai_credits from audit_snapshot.json
   - `ai_credits_trend.png`: dual-axis line chart from rolling-summary.json — total_ai_credits (left) and active_workflows (right, label "Active workflows/day"). Skip if fewer than 2 data points.
2. Upload each chart with upload_asset. Return the two URLs.

The smaller model handles Python script generation and execution well; the main agent retains responsibility for issue composition (Phase 4) where cross-referencing and synthesizing the full audit is required.

Expected additional savings from sub-agent: ~8–12 AIC/run (smaller model token cost for chart phase vs main model).

---

⚠️ Caveats

• Token breakdown is only available for 1 of 5 runs; per-phase token attribution is estimated from prompt structure analysis, not observed tool call logs.
• Avg 3 turns/run indicates the agent already completes efficiently — no turn-count reduction opportunities were identified.
• All 5 runs concluded success; no error-recovery paths were observed, so there is no reliability waste to eliminate.
• Sub-agent recommendation assumes the ## agent: inline sub-agent syntax is supported by this runner version.

Total estimated savings: 40–65 AIC/run (22–35% reduction)

Generated by Agentic Workflow AIC Usage Optimizer · 231.8 AIC · ⊞ 21.6K · ◷

• expires on Aug 11, 2026, 3:18 PM UTC