[ Protocol Draft · 01 ]// honest case for compression

The compression layer for agent communication.

AACP reduces agent-to-agent coordination tokens by 91.4% (measured, payroll workflow, 3 runs, Claude Sonnet 4.5). It does not reduce task tokens. Here is the honest case for why that still matters.

Slow · english47 tokens

"Please retrieve the employee salary
records for August 2024. I need all
active employees, their departments,
cost centres, base salary, deltas..."

Fast · aacp9 tokens · -81%

[TASK:FETCH][DOM:HR][RES:emp_salary][PERIOD:2024-08]
[FILTER:status=active][FIELDS:id,dept,cc,base_sal,delta,pension_rate]
[FMT:json][RETURN:HR-Agent][AACP:0.1]

91.4%

Coordination tokens · measured

Task token reduction (be explicit)

-5.6%

Cost · payroll workflow, cold-start, 3-run mean

-10.6%

Latency · payroll workflow, cold-start, 3-run mean

Source: payroll workflow · Claude Sonnet 4.5 · 3 runs · cold-start agent configuration.

Total cost to validate this protocol against a live API: $0.75

All measured benchmarks on this page combined · Claude Sonnet 4.5 · live API spend.

03 · Benchmarks

Measured — live API

Token counts come from live API usage_metadata. These are measured runs, not estimates. Task tokens are unchanged by AACP — only coordination tokens are reduced.

Measured · live API

Consistency callout

91% coordination token reduction measured consistently across both workflow types tested.

Payroll: -91.4% · IT Provisioning: -91.1% · both Claude Sonnet 4.5, 3 runs each.

Honest note · Coordination tokens drop ~91% across runs. Task tokens vary by model output and dominate total cost, so total cost / latency deltas swing run-to-run. AACP does not compress task work.

Workflow	Model	Runs	Hops	Coord tokens Δ	Total cost Δ	Latency Δ	Source	Timestamp
HR Payroll Run	claude-sonnet-4-5	3	4	-91.4%	-5.6%	-10.6%	Measured — live API	2026-05-23 14:32:07 UTC
IT Access Provisioning	claude-sonnet-4-5	3	3	-91.1%	+26.0%	+14.2%	Measured — live API	2026-05-23 15:23:55 UTC
HR Payroll Run	claude-sonnet-4-5	3	4	-91.4%	+19.9%	+6.0%	Measured — live API	2026-05-23 14:10:44 UTC

HR Payroll Run3 runs · 4 hops · claude-sonnet-4-5

Slow coord

163

AACP coord

-91.4%

Slow task

470

AACP task

879

+87.0%

Slow $

$0.0105

AACP $

$0.0099

-5.6%

Token counts are measured from live API usage_metadata. Accuracy scoring is based on Claude-evaluated semantic similarity. Results may vary across runs due to model non-determinism.

IT Access Provisioning3 runs · 3 hops · claude-sonnet-4-5

Slow coord

112

AACP coord

-91.1%

Slow task

282

AACP task

575

+103.9%

Slow $

$0.0042

AACP $

$0.0053

+26.0%

Token counts are measured from live API usage_metadata. Accuracy scoring is based on Claude-evaluated semantic similarity. Results may vary across runs due to model non-determinism.

Honest note · Cost increase reflects system prompt overhead in cold-start agent configuration. Coordination token reduction of 91.1% is consistent with payroll benchmark. Native persistent agents eliminate this overhead.

HR Payroll Run3 runs · 4 hops · claude-sonnet-4-5

Slow coord

163

AACP coord

-91.4%

Slow task

470

AACP task

1,165

+147.9%

Slow $

$0.0103

AACP $

$0.0123

+19.9%

Token counts are measured from live API usage_metadata. Accuracy scoring is based on Claude-evaluated semantic similarity. Results may vary across runs due to model non-determinism.

02 · Calculator

Run the numbers honestly

Honest framing

AACP compresses coordination tokens (routing, hand-off, metadata). It does not compress task tokens (the actual model work). Total workflow savings depend on the ratio between the two — typically 3–35% end-to-end, occasionally higher. Calculator figures below are projected from these ratios; see the Benchmarks section above for measured live-API runs.

Scenario

Volume / month

100

11001k10k

soleSMEmidcorp

Token breakdown · per workflow

6 hops × 2,800 task tok

Coordination tokens-91%

slow

1,428

aacp

126

Task tokensunchanged

slow

16,800

aacp

16,800

AACP does not compress task work

Total per run-7%

slow

18,228

aacp

16,926

	Without AACP	With AACP	Saving
Tokens / month	1,822,800	1,692,600	130,200
Cost / month	$5.47	$5.08	$0.39
Annual saving	—	—	$4.69

HR Payroll Run

FAVOURABLE

Monthly payroll across 6 agents: HR → Payroll → Tax → Bank → Audit → Notify.

Honest note · Best case. Repetitive structured data, clear task boundaries. Real savings are real here.

All scenarios · current model & volume

Scenario	Tag	Coord saving	Total saving	Monthly $
HR Payroll Run	FAVOURABLE	-91%	-7%	$0.39
Contract Review	REALISTIC	-90%	-1%	$0.19
Sales Lead Qualification	REALISTIC	-91%	-13%	$0.27
Open-Ended Research	STRESS TEST	-88%	-0%	$0.10
IT Access Provisioning	FAVOURABLE	-91%	-38%	$0.54
Customer Complaint	STRESS TEST	-86%	-4%	$0.17

Low total saving in task-heavy scenarios does not mean AACP fails — it means task cost dominates. Benchmark coordination cost separately.