Rate, Cost Limits and Audit Logs · ai-skills

Reference: Agentic Workflows for Chip Teams · Beginner to Intermediate

On this page

Rate and cost limits

Limits bound how much an agent may attempt. Rate limits cap actions per hour or day, and cost limits cap billed units per run.

Both are EXAMPLE parameters set before the run. The checker refuses new actions once a cap is reached.

Limit	EXAMPLE setting	What it stops
Steps per run	At most 30 tool calls	Loops that never converge.
Jobs per day	At most 20 simulation jobs	Queue flooding across a team.
Billed units per run	At most 8 cost units	An expensive retry spiral.

Worked cost arithmetic

Cost follows one formula with stated EXAMPLE prices. Substitute the run counts to get the run total in cost units.

plaintext

EXAMPLE parameters: 1M input tokens cost 3 units, 1M output tokens cost 12 units
formula: cost = input_tokens / 1M x input_price + output_tokens / 1M x output_price
per step: 40000 input tokens, 8000 output tokens
input part = 40000 / 1000000 x 3 = 0.04 x 3 = 0.12 units
output part = 8000 / 1000000 x 12 = 0.008 x 12 = 0.096 units
per-step total = 0.12 + 0.096 = 0.216 units
30-step run total = 30 x 0.216 = 6.48 units
check against EXAMPLE cap of 8 units: 6.48 < 8, so the run fits

plaintext

EXAMPLE daily budget: 20 simulation jobs per team per day
each triage run uses 2 jobs
runs per day = 20 / 2 = 10 runs
a proposed 11th run needs a wider budget or waits for the next day

Audit logs

The log records every proposed action, its verdict, and any approval. It is append-only and written by the checker, not the agent.

Each entry names who, what, and when. A minimal entry holds time, run id, action, target, verdict, and approver.

Field	Example value	Use
Time	2026-05-01 09:14 UTC	Orders events during review.
Run	triage-042	Groups one run's actions.
Action	repo write `rtl/alu.sv`	Names the tier and target.
Verdict	allow, needs-approval, deny	Shows the checker outcome.
Approver	named reviewer or none	Links each gate to a person.

Common mistakes

Check yourself The EXAMPLE run above costs 6.48 units against an 8-unit cap. How many more identical steps fit?

Each step costs 0.216 units, and 8 minus 6.48 is 1.52. Seven steps cost 1.512 and fit; eight cost 1.728 and do not.

Check yourself A log shows allows but no denials. What is the likely gap?

Denied proposals are missing. The agent writes drafts while the checker writes verdicts, so a complete log shows every verdict.

Related topics

Up nextSecret Handling

Practice this lesson Report an error

Rate and cost limits

Limits bound how much an agent may attempt. Rate limits cap actions per hour or day, and cost limits cap billed units per run.

Both are EXAMPLE parameters set before the run. The checker refuses new actions once a cap is reached.

Limit	EXAMPLE setting	What it stops
Steps per run	At most 30 tool calls	Loops that never converge.
Jobs per day	At most 20 simulation jobs	Queue flooding across a team.
Billed units per run	At most 8 cost units	An expensive retry spiral.

Worked cost arithmetic

Cost follows one formula with stated EXAMPLE prices. Substitute the run counts to get the run total in cost units.

plaintext

EXAMPLE parameters: 1M input tokens cost 3 units, 1M output tokens cost 12 units
formula: cost = input_tokens / 1M x input_price + output_tokens / 1M x output_price
per step: 40000 input tokens, 8000 output tokens
input part = 40000 / 1000000 x 3 = 0.04 x 3 = 0.12 units
output part = 8000 / 1000000 x 12 = 0.008 x 12 = 0.096 units
per-step total = 0.12 + 0.096 = 0.216 units
30-step run total = 30 x 0.216 = 6.48 units
check against EXAMPLE cap of 8 units: 6.48 < 8, so the run fits

plaintext

EXAMPLE daily budget: 20 simulation jobs per team per day
each triage run uses 2 jobs
runs per day = 20 / 2 = 10 runs
a proposed 11th run needs a wider budget or waits for the next day

Audit logs

The log records every proposed action, its verdict, and any approval. It is append-only and written by the checker, not the agent.

Each entry names who, what, and when. A minimal entry holds time, run id, action, target, verdict, and approver.

Field	Example value	Use
Time	2026-05-01 09:14 UTC	Orders events during review.
Run	triage-042	Groups one run's actions.
Action	repo write `rtl/alu.sv`	Names the tier and target.
Verdict	allow, needs-approval, deny	Shows the checker outcome.
Approver	named reviewer or none	Links each gate to a person.

Common mistakes

Check yourself The EXAMPLE run above costs 6.48 units against an 8-unit cap. How many more identical steps fit?

Each step costs 0.216 units, and 8 minus 6.48 is 1.52. Seven steps cost 1.512 and fit; eight cost 1.728 and do not.

Check yourself A log shows allows but no denials. What is the likely gap?

Denied proposals are missing. The agent writes drafts while the checker writes verdicts, so a complete log shows every verdict.