Target Workflow
Agentic Workflow AIC Usage Optimizer (agentic-token-optimizer)
Selected as the highest-AIC workflow not optimized in the last 14 days (last optimized 2026-08-20, 8 days ago). Second-highest total AIC in the 7-day window at 1,054 AI credits across 4 runs.
Spend Profile
| Metric |
Value |
| Analysis period |
2026-08-24 → 2026-08-27 |
| Completed runs |
4 |
| Total AIC |
1,054 |
| Avg AIC / run |
264 |
| Token data |
Not captured (null in run log) |
| Avg duration |
7–9 min / run |
| GitHub API calls / run |
10 (consistent) |
| Error rate |
0% (all 4 runs succeeded) |
| AIC variability |
High — range 163–328 (1.9× spread) |
Per-run AIC breakdown
The 163-AIC outlier (2026-08-26) suggests the agent selected a simpler target workflow and/or benefited from prompt caching that day.
Recommendations
1. Trim prompt verbosity — ~40–55 AIC/run
Action: Condense the three heavily-documented instructional sections in the prompt body.
The prompt is ~470 lines long. Of those, roughly 150–170 lines are instructional scaffolding that repeats on every run regardless of which workflow is selected:
- Data Access Guidelines (
## Data Access Guidelines): Contains 4 annotated bash code blocks with ✅/❌ annotations and inline comments. These patterns are already well-established — compress to 2–3 key bullet points and a single inline example.
- Inline Sub-Agent Scoring Rubric (
### Inline Sub-Agent Opportunity Analysis): Includes a 4-row scoring table, a 3-tier guidance block, two bullet lists ("Smaller models are a good fit for..." and "Keep with the main agent when..."), and a word limit reminder. This can be halved without loss of fidelity.
- Phase 3 workflow-source bash examples: Three separate bash snippets (full read, frontmatter-only, prompt-body-only) are embedded as documentation. The agent should use the most appropriate one; it doesn't need all three on every run.
Estimated savings: 150–170 lines × 2–3 tokens/line → ~400–500 tokens of input overhead per run. At the observed AIC/token ratio, that translates to ~15–20% AIC reduction ≈ 40–55 AIC/run.
Evidence: The 163-AIC run (2026-08-26) — the lowest-cost run — also had the shortest duration (6.2 min), suggesting that prompt-read overhead and analysis scope are the primary cost levers, not reliability or retries.
2. Inline sub-agent for workflow source analysis — ~25–40 AIC/run
Action: Extract Phase 3 + Phase 4 structural checks into an inline sub-agent using a smaller model.
Reading the target workflow's .md source and classifying its structure (configured tools, whether it has ## agent: blocks, major section headers, repeated setup prefixes) is entirely extractive and independent of the main agent's strategic reasoning. It is a strong candidate for a smaller model.
Score breakdown:
| Dimension |
Score |
Rationale |
| Independence |
3/3 |
Reads a fixed file path; needs no other phase outputs |
| Small-model adequacy |
3/3 |
File read → field extraction → classification; no synthesis needed |
| Parallelism |
2/2 |
Can run while the main agent finishes Phase 1/2 data aggregation |
| Size |
2/2 |
470-line source + structured output; substantial but bounded |
| Total |
10/10 |
Strong candidate |
Proposed change: Add an ## agent: block after Phase 2 with a concise prompt:
## agent: analyze-workflow-source
model: small
tools: [bash]
Read the workflow source at the path provided in `$TARGET_WF_PATH` using:
gh api "repos/$REPO/contents/$TARGET_WF_PATH" --jq '.content' | base64 -d
Extract and output as JSON:
- tools: list of configured tool names (from frontmatter)
- has_subagents: boolean (whether any "## agent:" block exists in the prompt body)
- major_sections: list of "##"-level section titles in the prompt body
- repeated_setup_lines: any lines appearing verbatim in 2+ section openings
The main agent then consumes this structured output instead of making raw gh api calls and parsing the source itself. This offloads ~80–120 lines of iterative file-reading and analysis work to a cheaper model call.
Estimated savings: ~25–40 AIC/run based on the sub-task's share of total run cost.
3. Batch file reads in Phase 1 — ~5–10 AIC/run
Action: Replace sequential reads of top-workflows.json, optimization-log.json, and optionally a daily snapshot into a single parallel tool call block.
Currently the agent reads these files separately across Phase 1 steps, adding multiple turn-round-trips. Issuing all three reads in one parallel block eliminates the intermediate inference steps between them.
Proposed change: At the start of Phase 1, consolidate:
# Current — 3 separate tool calls
cat /tmp/gh-aw/token-audit/top-workflows.json
cat /tmp/gh-aw/repo-memory/default/optimization-log.json
cat /tmp/gh-aw/repo-memory/default/$(date +%Y-%m-%d).json
into a single bash call:
jq -s '{"top": .[0].top_workflows, "log": .[1]}' \
/tmp/gh-aw/token-audit/top-workflows.json \
/tmp/gh-aw/repo-memory/default/optimization-log.jsonThis reduces Phase 1 turn count by 1–2 and eliminates redundant context re-reads.
Estimated savings: ~5–10 AIC/run.
Tool Usage Audit
| Tool |
Usage |
Status |
bash |
Used for all data reads, gh api calls, jq processing |
Keep |
github (issues toolset) |
Used to publish optimization issue via safeoutputs |
Keep |
repo-memory |
Read optimization-log.json, write updated log |
Keep |
All configured tools are used in every successful run. No removal candidates.
Caveats
- Token counts are
null in run logs for this workflow; AIC-based estimates assume proportional token scaling. Savings ranges are conservative (low end) estimates.
- The 163-AIC outlier (2026-08-26) may reflect an unusually short target workflow or prompt-cache hit; excluding it raises the average to 297 AIC/run and increases the savings opportunity.
- Inline sub-agent savings depend on the model cost differential between main and sub-agent models; if a small model is not available in this environment, the savings may be lower.
- Sample size: 4 completed runs. Recommendations should be validated across 5+ runs post-change.
References:
Generated by Agentic Workflow AIC Usage Optimizer · 275.3 AIC · ⊞ 21.6K · ◷
Target Workflow
Agentic Workflow AIC Usage Optimizer (
agentic-token-optimizer)Selected as the highest-AIC workflow not optimized in the last 14 days (last optimized 2026-08-20, 8 days ago). Second-highest total AIC in the 7-day window at 1,054 AI credits across 4 runs.
Spend Profile
Per-run AIC breakdown
The 163-AIC outlier (2026-08-26) suggests the agent selected a simpler target workflow and/or benefited from prompt caching that day.
Recommendations
1. Trim prompt verbosity — ~40–55 AIC/run
Action: Condense the three heavily-documented instructional sections in the prompt body.
The prompt is ~470 lines long. Of those, roughly 150–170 lines are instructional scaffolding that repeats on every run regardless of which workflow is selected:
## Data Access Guidelines): Contains 4 annotated bash code blocks with✅/❌annotations and inline comments. These patterns are already well-established — compress to 2–3 key bullet points and a single inline example.### Inline Sub-Agent Opportunity Analysis): Includes a 4-row scoring table, a 3-tier guidance block, two bullet lists ("Smaller models are a good fit for..." and "Keep with the main agent when..."), and a word limit reminder. This can be halved without loss of fidelity.Estimated savings: 150–170 lines × 2–3 tokens/line → ~400–500 tokens of input overhead per run. At the observed AIC/token ratio, that translates to ~15–20% AIC reduction ≈ 40–55 AIC/run.
Evidence: The 163-AIC run (2026-08-26) — the lowest-cost run — also had the shortest duration (6.2 min), suggesting that prompt-read overhead and analysis scope are the primary cost levers, not reliability or retries.
2. Inline sub-agent for workflow source analysis — ~25–40 AIC/run
Action: Extract Phase 3 + Phase 4 structural checks into an inline sub-agent using a smaller model.
Reading the target workflow's
.mdsource and classifying its structure (configured tools, whether it has## agent:blocks, major section headers, repeated setup prefixes) is entirely extractive and independent of the main agent's strategic reasoning. It is a strong candidate for a smaller model.Score breakdown:
Proposed change: Add an
## agent:block after Phase 2 with a concise prompt:The main agent then consumes this structured output instead of making raw
gh apicalls and parsing the source itself. This offloads ~80–120 lines of iterative file-reading and analysis work to a cheaper model call.Estimated savings: ~25–40 AIC/run based on the sub-task's share of total run cost.
3. Batch file reads in Phase 1 — ~5–10 AIC/run
Action: Replace sequential reads of
top-workflows.json,optimization-log.json, and optionally a daily snapshot into a single parallel tool call block.Currently the agent reads these files separately across Phase 1 steps, adding multiple turn-round-trips. Issuing all three reads in one parallel block eliminates the intermediate inference steps between them.
Proposed change: At the start of Phase 1, consolidate:
into a single bash call:
jq -s '{"top": .[0].top_workflows, "log": .[1]}' \ /tmp/gh-aw/token-audit/top-workflows.json \ /tmp/gh-aw/repo-memory/default/optimization-log.jsonThis reduces Phase 1 turn count by 1–2 and eliminates redundant context re-reads.
Estimated savings: ~5–10 AIC/run.
Tool Usage Audit
bashgithub(issues toolset)repo-memoryoptimization-log.json, write updated logAll configured tools are used in every successful run. No removal candidates.
Caveats
nullin run logs for this workflow; AIC-based estimates assume proportional token scaling. Savings ranges are conservative (low end) estimates.References: