[agentic-token-optimizer] Daily Agentic Workflow AIC Usage Audit — Optimization Recommendations

[github-actions]

🎯 Target Workflow

Daily Agentic Workflow AIC Usage Audit (agentic-token-audit)
Selected as the highest-AIC workflow in the 7-day window (1,117.78 AIC across 5 runs). All workflows in this repository have a recent optimization entry; this one has the highest ongoing spend and a new run today (run #60) that shows a notable duration spike useful for evidence.

Note: All five tracked workflows fall within the 14-day recency window. This workflow was last addressed on 2026-07-22 (5 days ago); the recommendations below focus on newly observable patterns from today's run and areas not previously targeted.

---

📊 Spend Profile (7-day window, 2026-07-21 → 2026-07-27)

Metric	Value
Runs analyzed	5
Total AIC	1,117.78
Avg AIC / run	223.56
Total tokens (1 run with data)	606,644
Avg turns / run (1 run with data)	16
Total Actions minutes	49 min
All conclusions	✅ success × 5

Per-run breakdown

Run	AIC	Tokens	Turns	Duration
§29831796021	200.06	—	—	8.0 min
§29921394771	262.46	—	—	7.8 min
§30008781105	281.59	—	—	8.1 min
§30094628596	167.53	—	—	7.1 min
§30270012569	206.15	606,644	16	15.5 min

Token/turn data only available for run #60 (today's run). Duration spike on run #60 (15.5 min vs 7–8 min average) is a clear regression signal.

---

🔍 Analysis

Tool Usage

The workflow configures agentic-workflows, bash, and repo-memory. All are used legitimately across runs. GitHub API calls per run are 8–10, consistent with reading workflow sources in the setup step and writing repo-memory. No tools appear configured but unused.

AI Credit Spend Drivers

1. Large prompt body (333 lines): includes a 9-field data schema table, inline JSON schema examples for two output files, detailed chart rendering instructions, and a conditional OTEL span section — all loaded into context on every turn.
2. Write-then-run script pattern (Phase 1): the agent is instructed to write a Python script to /tmp/gh-aw/token-audit/process_audit.py and then execute it. This is a multi-turn sequence (write file → confirm write → run → read output) that adds 2–3 overhead turns.
3. Chart upload / URL placeholder substitution (Phase 3): upload two assets, receive URLs, then craft the issue with embedded URLs — another multi-turn sequence.
4. Duration spike on run Consolidate optimizer workflows into a single token optimizer #60 (15.5 min): 7 more minutes than the 4-run average suggests one of the above sequences expanded noticeably.

Reliability

All 5 runs succeeded. No error or warning patterns. No waste from retries.

---

📋 Ranked Recommendations

1. Trim the data schema table to essential fields only

Estimated savings: ~25 AIC/run

The prompt's "Pre-downloaded logs" section includes a 9-row Markdown table describing every field of a RunData object — including deprecated (effective_tokens) and rarely-used fields (error_count, warning_count, token_usage_summary). The agent's Python script only needs workflow_name, status, conclusion, aic, token_usage, turns, action_minutes, url, and run_id. The OTEL section also has a Note column and deprecation warnings adding noise.

Action: Reduce the schema table to the 9 fields actually used in Phase 1's aggregations, and remove the "effective_tokens — deprecated; use aic" row entirely. Estimated ~50-line reduction in the prompt body.

2. Condense or extract the OTEL experiment section

Estimated savings: ~12 AIC/run

The "Experiment OTEL Span Attributes" section is ~25 lines of JavaScript boilerplate that only activates when /tmp/gh-aw/experiments/assignments.json exists. It adds context overhead on every run regardless of whether experiments are active. It also includes a full code block that the agent must process on each turn.

Action: Reduce to a two-sentence conditional instruction: "If /tmp/gh-aw/experiments/assignments.json exists, call the logSpan OTEL helper (see the shared component for the pattern) to emit one gh_aw.experiment.<name> attribute per experiment key." If the full code block must be retained, move it to a shared component imported by reference.

3. Replace the "write script to file then run" pattern with inline Python execution

Estimated savings: ~15 AIC/run

Phase 1 tells the agent to write a Python script to a named file and then execute it with a custom PYTHONPATH. Writing a file to disk, confirming the write, and then running it takes 2–3 extra turns compared to executing the Python logic inline via python3 -c '...' or a heredoc. The computation (grouping, aggregation, JSON output) is simple enough for either approach.

Action: Change Phase 1's instruction from "write a script to process_audit.py and run it" to "run the data processing inline using python3 << 'EOF' ... EOF with the PYTHONPATH prefix." Retain the schema and sort requirements in the instruction; just eliminate the write-to-file indirection. For the chart generation in Phase 3, a file script is still appropriate due to length.

---

🧩 Structural Optimization: Inline Sub-Agent for Log Processing

The workflow has no existing ## agent: blocks and has 4 distinct phases.

Phase 1 (Process Logs) as an inline sub-agent:

Dimension	Score	Reasoning
Independence	3/3	Reads only pre-downloaded files; no dependency on other phases
Small-model adequacy	3/3	Pure aggregation: group-by, sum, sort, serialize JSON
Parallelism	0/2	Phases 2, 3, 4 all depend on its output
Size	2/2	Writes a Python script + runs it; substantial enough
Total	8/10	Strong candidate

Proposed change:

In the main prompt, replace the Phase 1 block with:

## agent: process-logs
model: small

Read `/tmp/gh-aw/token-audit/workflow-logs.json`. Filter to `status == "completed"` runs. Group by `workflow_name` and compute: `run_count`, `total_ai_credits`, `avg_ai_credits`, `total_tokens`, `avg_tokens`, `total_turns`, `avg_turns`, `total_action_minutes`, `error_count`, `warning_count`. Treat null/missing `aic`/`token_usage` as 0. Sort descending by `total_ai_credits`. Write the result plus an overall summary to `/tmp/gh-aw/token-audit/audit_snapshot.json`.

Why a smaller model fits: This is pure ETL — load, aggregate, serialize. No strategic reasoning, no synthesis, no GitHub API calls. A smaller model handles Python aggregation reliably.

Expected savings from sub-agent: ~20 AIC/run (smaller model, fewer context tokens in that phase).

---

💰 Combined Estimated Savings

Recommendation	Est. AIC / run
Trim schema table	~25
Condense OTEL section	~12
Inline Python execution (Phase 1)	~15
Sub-agent for log processing	~20
Total	~72 AIC/run

At 5 runs/week × ~223 AIC/run current spend, implementing all four could reduce weekly spend by ~32% (~360 AIC/week).

---

⚠️ Caveats

• Token data is only available for 1 of 5 runs (run Consolidate optimizer workflows into a single token optimizer #60). Savings estimates are extrapolated from that single run's 606,644 token count and 16 turns.
• The duration spike on run Consolidate optimizer workflows into a single token optimizer #60 (15.5 min) may reflect a one-time condition (e.g., larger rolling-summary file, slower chart rendering). Monitor the next 3–5 runs before attributing it to prompt length.
• The sub-agent recommendation requires the gh-aw runtime to support ## agent: with model: small inline blocks. Verify this feature is available before implementing.
• All four recommendations can be applied independently; start with Add CI workflow to compile agentic workflows via gh-aw-actions #1 and Update ci.yml: add workflow_dispatch trigger and add job #2 (purely editorial) as they carry no behavioral risk.

References: §30270012569 · §30008781105 · §30094628596

Generated by Agentic Workflow AIC Usage Optimizer · 154 AIC · ⊞ 21.6K · ◷

• expires on Aug 3, 2026, 3:28 PM UTC

[github-actions]

This issue is being closed as outdated. A newer issue has been created: #344

View newer issue

---

This action was performed automatically by the Agentic Workflow AIC Usage Optimizer workflow.