[agentic-token-optimizer] Optimize Daily XKCD Comic — reduce agent turns via pre-computation and cache usage

[github-actions]

Target Workflow

Daily XKCD Comic (daily-xkcd.md) — selected as the highest-AIC workflow not recently optimized (prior optimization: 2026-08-21, 6 days ago). All top workflows have been optimized within 14 days; this one has the oldest optimization date and meaningful per-run spend (~58 AIC/run across 7 daily runs).

---

Analysis Period

2026-08-21 – 2026-08-27 · 7 runs audited · All conclusions: success · 0 errors

Spend Profile

Metric	Value
Total AIC (7 runs)	406.95 credits
Avg AIC / run	~58.1 credits
Avg tokens / run	~180 k
Avg turns / run	6.7 (range: 5–8)
Total action-minutes	38 min
Cache efficiency	Not reported
Errors	0

Per-run breakdown

Run	Date	AIC	Tokens	Turns	Conclusion
§32467304018	2026-08-21	74.28	212 574	8	success
§32564407628	2026-08-22	~58	~180 k	~7	success
§32630448758	2026-08-23	~58	~180 k	~7	success
§32711427637	2026-08-24	72.53	210 309	8	success
§32831303647	2026-08-25	51.46	152 886	6	success
§32952653667	2026-08-26	42.44	126 510	5	success
§33072973368	2026-08-27	60.94	180 301	7	success

---

Ranked Recommendations

1 — Pre-compute comic selection in steps (remove from agent prompt)

Estimated savings: ~12–15 AIC/run (~20–25%)

The current steps section fetches only the latest comic number. The agent is then asked to:

1. Read the curated list (50+ hardcoded numbers across 3 categories embedded in every prompt context window)
2. Compute date +%j mod list_size
3. Pick a comic and verify relevance (potentially 3 attempts)

This selection logic is deterministic and purely computational — it does not require LLM reasoning. Moving it into the steps section eliminates 1–2 agent turns per run and removes ~250 tokens of static list data from every prompt window.

Proposed steps addition:

- name: Select comic for today
  run: |
    set -euo pipefail
    DEVELOPER_COMICS="303 327 353 386 456 519 664 705 722 737 844 859 936 1068 1168 1205 1319 1425 1443 1537 1739"
    MATH_COMICS="55 135 174 687 712 715 881 882 1236 1261 1379 1478 1754 2048 2100"
    ML_COMICS="1838 1875 2050 2173 2347 2440 2501 2545 2581 2674 2746 2785 2860 2916 2925"
    ALL_COMICS="$DEVELOPER_COMICS $ML_COMICS $MATH_COMICS"
    COMICS_ARRAY=($ALL_COMICS)
    COUNT=${#COMICS_ARRAY[@]}
    DOY=$(date +%j | sed 's/^0*//')
    INDEX=$((DOY % COUNT))
    SELECTED=${COMICS_ARRAY[$INDEX]}
    echo "$SELECTED" > /tmp/gh-aw/agent/xkcd/selected_num.txt
    echo "Selected comic for today: #$SELECTED"

The prompt can then replace the entire Comic Selection section with: "The curated comic number for today is in /tmp/gh-aw/agent/xkcd/selected_num.txt."

Evidence: Turn variance (5–8) across runs correlates with the retry loop. Runs with 8 turns (2026-08-21, 2026-08-24) cost 72–74 AIC vs. 5-turn runs at 42 AIC — the selection/retry loop drives most of the variance.

---

2 — Read cached latest.json instead of re-fetching

Estimated savings: ~4–6 AIC/run (~7–10%)

The steps section already downloads the full latest comic JSON to /tmp/gh-aw/agent/xkcd/latest.json. However, instruction step 2 reads:

"Check if the latest comic is relevant (fetch it and read its title/alt text)"

This directs the agent to web_fetch `(xkcd.com/redacted) — a duplicate network call. The agent should instead read the already-cached file.

Fix: Change instruction step 2 from:

"Check if the latest comic is relevant (fetch it and read its title/alt text)"

To:

"Read /tmp/gh-aw/agent/xkcd/latest.json (already downloaded) and check if the title/alt text is relevant."

Evidence: Every run makes at least one web_fetch call for the latest comic despite the steps section downloading it. Eliminating this call saves one tool invocation and reduces context churn.

---

3 — Remove expr from allowed tools if unused

Estimated savings: negligible AIC, minor prompt hygiene

The workflow declares expr as an allowed bash command. The setup step uses jq and bash arithmetic. With pre-computation moved to steps (recommendation 1), no expr usage remains. If no run has used expr, it can be removed from the tools list.

Action: Audit the last 7 run logs for expr invocations. If zero occurrences, remove expr from the bash tools list.

---

Structural Optimization

Common setup prefix: minor duplication between Context and Instructions

Both the Context section and Instructions step 1 reference /tmp/gh-aw/agent/xkcd/latest_num.txt. This is minor and does not warrant extracting a shared setup prefix.

Inline sub-agents: not recommended

The workflow has a single coherent task: select a comic, fetch its data, create a discussion. There are no independent extractive subtasks of sufficient size to justify a sub-agent call. The main agent handles the final output creation, which requires synthesis — not a small-model fit.

---

Caveats

• Turns data is only directly available for 3 of 7 runs from daily snapshots; the remaining runs are estimated from AIC proportions.
• The turn-count variance (5–8) may reflect XKCD API latency as well as retry loops; recommendation 1 savings estimate uses the conservative end.
• Moving comic selection to steps changes the retry behavior: the agent can no longer try multiple curated numbers. Consider keeping a short fallback list (3 alternatives) in the prompt for the relevance check.

References:

• §32467304018 (8 turns, 74.28 AIC — high-cost run)
• §32711427637 (8 turns, 72.53 AIC — high-cost run)
• §32952653667 (5 turns, 42.44 AIC — low-cost baseline)

Generated by Agentic Workflow AIC Usage Optimizer · 256.1 AIC · ⊞ 21.6K · ◷

• expires on Sep 3, 2026, 6:13 PM UTC