[agentic-token-optimizer] Daily Agentic AI Research Digest — Reduce over-fetching and tighten browsing cap

[github-actions]

Target Workflow

Daily Agentic AI Research Digest (daily-agentic-research.md)
Selected as highest-AIC non-monitoring workflow not excluded by the 14-day cooldown. All other top workflows were optimized within the last 7 days.

Analysis Period

14 runs (2026-07-09 → 2026-07-28) from daily snapshots + all-runs.json.

Spend Profile

Metric	Value
Total AIC (14 runs)	871.73
Avg AIC / run	62.27
Total tokens	2,550,226
Avg tokens / run	182,159
Avg turns / run	6.7
Turn range	4 – 12
High-turn runs (≥ 8 turns)	4 of 14 (29%)
Error rate	0% (all 14 succeeded)
Cache efficiency	Low — fresh web pages fetched daily, no caching benefit

Avg ~27,100 tokens per turn; a single over-fetched page adds ~25–30K tokens (~8–10 AIC).

Ranked Recommendations

1. Add a hard fetch limit — Estimated savings: 10–20 AIC / run

Evidence: 4 of 14 runs consumed ≥ 8 turns (up to 12), spending 2–3× the AIC of efficient 4-turn runs. The current instruction is soft: "Browse 2–3 of these sources (don't fetch all if you find a strong candidate early)" — but the agent still fetches 3–4 sources on ~29% of runs.

Action: Replace the current soft guidance in Browsing Instructions with an explicit hard cap:

Fetch at most 2 sources. Stop as soon as you have identified a candidate
that meets the Selection Criteria — do not continue fetching.

Dropping just 1 extra web-fetch (~27K tokens) saves ~9 AIC/run on the 29% of runs that over-fetch, yielding ~2.6 AIC expected savings per run. Eliminating 2+ extra fetches on the worst-case runs (July 10: 12 turns, 111 AIC) can save 15–20 AIC on those runs alone.

Conservative estimate: 10–15 AIC/run average across all runs.

---

2. Trim the source list from 4 to 2 primary sources — Estimated savings: 3–8 AIC / run

Evidence: The four configured sources are huggingface.co/papers, arxiv.org, openai.com/news, and anthropic.com/news. OpenAI and Anthropic news pages publish product announcements more than research findings, and the selection criteria explicitly favor "specific and concrete" agentic research. Having 4 candidate sources in scope drives the agent to attempt more fetches.

Action: Remove openai.com/news and anthropic.com/news from Browsing Instructions. Keep huggingface.co/papers and arxiv.org as the two primary research-paper sources. These two are sufficient to surface daily agentic AI research.

Conservative estimate: 3–8 AIC/run by reducing average sources fetched.

---

3. Tighten the fallback instruction — Estimated savings: 2–5 AIC / run on fallback runs

Evidence: The current fallback — "If nothing brand-new is available, pick the most underappreciated or actionable recent finding" — encourages the agent to scan more broadly when it doesn't find something from the last 3 days, driving additional fetches.

Action: Add a strict time fence to the fallback:

If nothing from the last 7 days is found in the first 2 sources, stop —
do not fetch additional sources. Use the best item found regardless of age.

This prevents the 3-fetch cascade that the current open-ended fallback can trigger.

Conservative estimate: 2–5 AIC/run when fallback path is taken.

---

Run evidence — all 14 runs

Date	AIC	Tokens	Turns
2026-07-09	47.24	135,687	5
2026-07-10	111.15	328,672	12
2026-07-13	46.70	136,505	5
2026-07-14	62.83	184,635	7
2026-07-15	92.07	269,982	10
2026-07-16	75.04	219,744	8
2026-07-17	36.98	106,930	4
2026-07-20	89.21	257,505	9
2026-07-21	55.35	161,660	6
2026-07-22	44.78	132,202	5
2026-07-23	65.37	191,267	7
2026-07-24	46.03	133,781	5
2026-07-27	43.56	129,198	5
2026-07-28	55.42	162,458	6

High-turn runs (≥ 8): July 10, 15, 16, 20 — all clustered around broader browsing sessions.

References:

• §30350694237 — 2026-07-28, 6 turns, 55 AIC (typical)
• §30259543920 — 2026-07-27, 5 turns, 43 AIC (efficient)

Caveats

• The July 10 run (12 turns, 111 AIC) is an outlier that may inflate the high-turn average; with it excluded, average turns drop to 6.2.
• Only snapshot-level data is available for earlier runs (no per-turn tool logs), so exact fetch counts are inferred from turn counts.
• Sub-agent refactoring is not recommended: the workflow is compact (one tool, one output), and adding sub-agent overhead would likely increase total AIC rather than reduce it.
• Tool configuration is already minimal (web-fetch only); no tool removal opportunities.

Generated by Agentic Workflow AIC Usage Optimizer · 154.1 AIC · ⊞ 21.6K · ◷

• expires on Aug 4, 2026, 3:15 PM UTC