GAP-07 · Agent architecture

Protocol Selection Lab

See why a protocol leads and which input would change the recommendation.

Current decision

MCP leads this decision model

One shared protocol for tool/resource access across many systems and consumers. This is the highest weighted score for the six answers currently selected; inspect the actual flip conditions below.

Leading score: 4.8
Runner-up: MCP + A2A hybrid
Score margin: 0.4

Same instrument · 3 industries choosing a scenario replaces its inputs; save or share your current changes first

Prefer to read? The two minute case study · problem → approach → metric → outcome

Problem

Different protocols optimize different problems. Function calling can be simpler for narrow workflows, MCP can standardize tools across consumers, A2A can support agent coordination, and hybrid approaches may be required when the enterprise environment has multiple integration surfaces.

Approach

The model scores a protocol recommendation across weighted criteria and shows the runner up, the sensitivity of the recommendation, and the condition that would flip the decision.

Why this way

This connects architecture choice to standardization, reusability, delivery speed, change cost, governance, and operating complexity.

The metric

The fit scores and the flip conditions: which single input change would change the recommendation.

The trade-off

Broadening scope raises value but lowers feasibility; the radar shows function calling and MCP as mirror opposites on tool breadth, you choose where on that curve to sit.

Outcome

A protocol recommendation you can defend in a design review, with the runner up and the exact condition that would flip it, exportable as a one page card.

What changed the recommendation?

Baseline and current inputs match. MCP remains first.

1MCP4.8 points0.0 from baseline

2MCP + A2A hybrid4.4 points0.0 from baseline

3Function calling3.3 points0.0 from baseline

4A2A3.2 points0.0 from baseline

Compare all scores with the baseline

Protocol score changes
Protocol	Baseline	Current
Function calling	3.3	3.3
MCP	4.8	4.8
A2A	3.2	3.2
MCP + A2A hybrid	4.4	4.4

Only real rank changes move the rows. Scores are model points, not probabilities or measured protocol performance.

How many systems / tools to expose?

How many agent consumers?

Coordination needs?

Governance / central control?

Reuse across teams?

Simplicity sensitivity?

Recommendation

MCP

One shared protocol for tool/resource access across many systems and consumers.

Many systems and many consumers: bespoke integrations explode as N×M. One MCP contract makes it N+M.

Runner up: MCP + A2A hybrid. Inspect the calculated single-answer flip conditions below; a general preference is not a guaranteed flip.

Why not the others

vs Function calling: MCP leads on Tools.
vs A2A: MCP leads on Tools.
vs MCP + A2A hybrid: MCP leads on Tools.

Why this call

MCP, a toss-up call, +0.4 ahead of MCP + A2A hybrid.
Your "Tools" answer points to MCP.
Closest alternative: MCP + A2A hybrid.
It would flip to Function calling if "How many systems / tools to expose?" became "1 to 3 tools".

Shareable card

Fit scores

Function calling3.3

MCP4.8

A2A3.2

MCP + A2A hybrid4.4

Protocol shapes · responsiveness to each decision dimension

Function callingMCPA2AMCP + A2A hybrid

Outer = the dimension favors that protocol; the dashed mid ring is neutral; inside it the dimension argues against. Moves with your weights.

Read protocol responsiveness values

Protocol responsiveness by decision dimension
Protocol	Tools	Consumers	Coordination	Governance	Reuse	Simplicity
Function calling	14.6 / 100	34.4 / 100	16.7 / 100	50.0 / 100	50.0 / 100	39.6 / 100
MCP	83.3 / 100	72.9 / 100	50.0 / 100	70.8 / 100	72.9 / 100	50.0 / 100
A2A	50.0 / 100	66.7 / 100	100.0 / 100	50.0 / 100	50.0 / 100	50.0 / 100
MCP + A2A hybrid	50.0 / 100	69.2 / 100	97.9 / 100	64.6 / 100	50.0 / 100	50.0 / 100

Producers × consumers

≈ 7 systems × 4 consumers28 bespoke11 protocol

Bespoke point to point integrations grow as producers×consumers; a shared protocol grows as producers+consumers. Past ~2 consumers the protocol wins, you're at 4.

What would change the call · sensitivity

How many systems / tools to expose?→ answer “1 to 3 tools” and it flips to Function calling
How many agent consumers?→ answer “One” and it flips to Function calling
Coordination needs?→ answer “One agent does it” and it flips to Function calling
Reuse across teams?→ answer “One-off” and it flips to MCP + A2A hybrid

If you act on this · the call → expected lift → how you'd measure it

The call

Select the protocol based on the integration pattern, not the trend cycle.

Expected lift · illustrative

Improves architecture consistency and avoids premature standardization or unnecessary custom integration.

How you'd measure it

Integration reuse, onboarding time, tool adoption, change effort, support complexity.

Why the runner up matters

A recommendation without a runner up is a quiz answer. The flip condition tells leadership exactly what would change the architecture, so the decision survives the next scale-up instead of being re-litigated.

Steering committee takeaway: The protocol is not the first question. Count the systems, consumers, coordination needs, and governance constraints. The protocol follows from that shape.

How this is built

Each protocol scores against the six answers: function calling rewards small surface + single consumer; MCP rewards systems × consumers + governance + reuse; A2A rewards multiagent coordination; hybrid tops only when MCP and A2A signals are both strong.

Primary = top score, runner up = second; the flip condition names the runner up's dominant driver. Protocol landscape stats are dated config (as of 2026-07-02).

Stack: Next.js (static) + shared design system; client side only.

Limitations: this artifact is a decision model, not a full architecture review. Production selection would require security review, platform standards, vendor constraints, latency requirements, and implementation testing.