|
Five signals, one direction:
| 1. 150+ organizations running A2A in production (April 2026 anniversary) |
| 2. Agent economy is real: Agentic commerce at $20.9B (1.5% of retail), early adopters seeing 6–10% revenue growth + 40% efficiency gains |
| 3. MCP + A2A = the operating system for enterprise agents: 97M MCP installs, 10K+ enterprise servers; A2A v1.2 landed March 2026 |
| 4. Pricing collapsed 60–80%: Token costs at historic lows ($0.25/M floor), shifting from per-seat to per-outcome billing |
| 5. Multi-vendor interop is now mandatory: No enterprise agent system ships without protocol alignment (discovery, security, task governance) |
The inflection point: A year ago, A2A was positioning. Today, it’s infrastructure. Salesforce agents talk to Google agents through standardized Agent Cards. Stripe processes A2A commerce transactions. Discovery is moving from “handshake integrations” to protocol-driven registries. This is the first time in enterprise AI history that vendor lock-in is optional.
⚙️ Part 1: A2A Architecture — What Actually Happened at Google
April 9, 2025, Google announced Agent-to-Agent (A2A) with 50+ founding partners. One year later (April 9, 2026), the protocol hit 150+ production organizations. That’s not gradual adoption — that’s tipping-point acceleration.
Think of MCP as “how agents talk to tools.” A2A is “how agents talk to agents.”
Google designed A2A around five principles:
| Principle |
What It Solves |
| Agentic Capabilities | Agents stay opaque; vendor lock-in eliminated |
| Build on Standards | HTTP/1.1, JSON-RPC 2.0, SSE, HTTPS — no ecosystem fragmentation |
| Security by Default | Signed Agent Cards, OAuth 2.0, mTLS, scoped tokens |
| Long-Running Tasks | Explicit task lifecycle (pending → in-progress → completed/failed) |
| Modality Agnostic | Text, audio, video, files, structured data in single task |
The Core Components:
| Agent Card (/.well-known/agent-card.json) — JSON metadata published by every A2A agent. Describes capabilities, endpoints, auth requirements. Discovery is zero-config: just fetch the card. Salesforce CRM agent advertises “query leads,” “create opportunity,” “update forecast.” |
| Task Management — Client agent creates a task with intent, context, constraints. Remote agent transitions it through defined states. Streaming updates via Server-Sent Events (SSE). |
| Message Exchange — Structured payloads with “parts” (text, artifacts, forms). Agents negotiate content types mid-conversation. Real-time push notifications on status changes. |
| Security Layer — OAuth2, OpenID Connect, API keys. Authorization verified by remote agent. Audit trails for compliance (healthcare, finance, legal). Zero-trust: each agent is an independent security boundary. |
Real Implementation: Tyson Foods + Gordon Food Service
Tyson and GFS deployed collaborative A2A systems in Q1 2026. Tyson’s inventory agent publishes availability and pricing. GFS’s demand planning agent queries Tyson, compares with competitors, negotiates bulk deals in real time. Result: 12–18% reduction in supply chain latency, $2.3M in margin recovery (Q1 2026). Before A2A, one new integration took 6 weeks. Post-A2A: 3 days.
Enterprise production signals: Microsoft Semantic Kernel (A2A support, Q1 2026) • Salesforce Agentforce (native A2A; Box, Workday, ServiceNow plug in directly) • AWS Bedrock AgentCore (A2A roadmap, Q2 2026) • Azure AI Foundry (A2A in production deployments)
📈 Part 2: A2A vs MCP — Why They’re Not Competing
This is the question that kills 40% of initial planning meetings. Here’s the answer:
| Protocol |
Direction |
Focus |
Adoption |
| MCP |
Agent → Tools (vertical) |
Context, tool discovery, permission scoping |
97M installs; foundational |
| A2A |
Agent ↔ Agent (horizontal) |
Task handoff, capability discovery, long-running orchestration |
150+ orgs in production; accelerating |
The Stack (visualized):
| A2A Layer |
Multi-agent workflow — agent-to-agent coordination, discovery, task management |
| MCP Layer |
Individual agent — tool access, context, external data (Salesforce, GitHub, Postgres) |
| Foundation Model |
Claude, GPT-5.5, Gemini 3.1 — reasoning and generation |
Real World (Capital One Finance): MCP layer lets the loan agent access core banking APIs, credit bureau feeds, internal data warehouse. A2A layer lets the loan agent delegate risk scoring to a Risk agent and approval to a Compliance agent. Result: 100% automated decisioning on $2.7B in Q1 2026 loan volume, zero manual escalation.
Why both? MCP alone = agent is smart but isolated. A2A alone = agents coordinate but can’t access external data. MCP + A2A = full-stack agentic systems (cross-vendor, multi-domain, governed). Google ADK now ships both out-of-box. Anthropic and OpenAI are founding members of A2A governance (Linux Foundation).
🏢 Part 3: Enterprise Implications — Discovery, Security, Governance
Three hard problems emerged once A2A hit scale:
|
Problem 1: Discovery at Scale
Fortune 500 firm, 59 agents total (47 internal + 12 vendor). Pre-A2A: hardcoded endpoints, spreadsheet of “who talks to whom,” nightmare to update. Post-A2A (AGNTCY Framework): Agent Cards in a registry, semantic search on capability, version negotiation baked in.
Cisco (AGNTCY Project): 200+ agents across 15 business units. Time to onboard new agent: 30 minutes (was 4 weeks). $18M in operational cost savings (2025). Open-source on GitHub.
|
|
Problem 2: Security Across Org Boundaries
Salesforce CRM agent (cloud) needs to query a supplier agent (on-prem). A2A solution: HTTPS + optional mTLS (transport), OAuth 2.0 (auth), scoped tokens (authorization), structured task IDs (audit).
Deloitte + large pharma (FDA 21 CFR Part 11, Q1 2026): Pre-A2A: custom SSL tunnels, RBAC hacks. Post-A2A: mutual TLS, signed Agent Cards, audit trail per task. FDA compliance verified Jan 2026. 12% faster batch releases.
|
|
Problem 3: Governance — Who Controls the Workflow?
Three agents on a $50M procurement deal: internal, vendor, finance. Who’s liable if vendor agent makes a bad recommendation? Can it see other vendors’ pricing?
FifthRow Governance Pattern (April 2026): Task boundaries scope what vendor can see. State isolation limits context per role. Finance agent injects policies before task transitions. Human-in-loop option for deals above $1M.
Unilever (Q1 2026): 150+ supplier negotiation agents. Human escalation: 12% → 4%. Cost per deal: $340 → $180 (47% reduction).
|
💰 Part 4: The Agent Economy — Marketplaces, Pricing, ROI
Market Size (Mordor Intelligence, April 2026): Agentic AI: $6.96B (2025) → $9.89B (2026) → $57.42B (2031). CAGR: 42%. Faster adoption than cloud compute (30% CAGR 2009–2015).
Eight Marketplaces That Matter (Q2 2026):
| Marketplace |
Model |
Use Case |
| Agentspace (Google Cloud) | Free listing; commission on enterprise deals | Enterprise, multi-vendor |
| ClawHub (OpenAI) | Per-task; creator takes 70% | Consumer agents |
| AgentExchange | Free publish; subscriber pays $9–$99/mo | Developers, open source |
| Anthropic Claude Marketplace | Per-outcome: $0.50 per resolved conversation | Customer service agents |
| Agentalent.ai (Monday.com) | “Hire” agents; $99–$999/mo | SMBs, no-code workflows |
The Pricing Shift:
| Pattern |
Example |
When It Works |
| Per-Outcome | HubSpot: $0.50 per resolved conversation | High-volume, measurable outcomes |
| Per-Consumption | Gemini Flash-Lite: $0.25/M tokens | Variable workload, scaling businesses |
| Hybrid (Seat + Usage) | Salesforce Agentforce: $500/mo + credits | Enterprise; predictable + variable |
The flat-fee era is over. Anthropic moved enterprise billing to per-token in April 2026. Every lab is expected to follow within six months.
Early Adopter ROI — Quantified:
| +6–10% revenue increase for early adopters (Rierino, Q1 2026; 200+ orgs) |
| 40% order efficiency improvement (fulfillment, supply chain) |
| 4–6 week payback period (vs 6–12 months for in-house builds) |
Mid-Market SaaS Case Study (Q1 2026): 3 agents (onboarding, subscription management, churn prediction). Investment: $180K + $4.2K/mo OpEx. Q1 result: +$220K incremental revenue. ROI: 122% in 90 days.
Shopify Agentic Storefronts (Jan 2026): Early adopters using agentic checkout see 2.1x higher conversion vs standard web flow. Agents now sell simultaneously across ChatGPT, Perplexity, Microsoft Copilot.
🔥 Weekly AI Roundup: April 14–20, 2026
|
1. Claude Opus 4.7 vs 4.6: Quality Plateau, Price Holds
Opus 4.7 wins 12 of 14 benchmarks; same $5/$25 pricing. Anthropic is competing on quality, not cost.
For builders: if you’re locked into Claude for reliability, you pay premium. For volume workloads, switch to Gemini Flash-Lite ($0.25/M).
|
|
2. Google Gemini 3.1 Flash-Lite: Pricing Floor at $0.25/M Tokens
2.5x faster, 45% cheaper output vs Flash. Sets new price floor. Per-token costs collapsed 60–80% in 12 months.
Model routing strategy wins: route simple tasks to Flash-Lite, complex to Opus. Blended cost = material savings.
|
|
3. Anthropic Enterprise Shifts to Per-Token Billing
Moved away from flat-rate contracts. OpenAI, Google expected to follow within 6 months. Industry standard shift confirmed.
Action required: implement cost caps, rate limits, token budgets now. Per-token surprise bills are a risk starting today.
|
|
4. A2A Protocol Hits 150 Organizations (April 9 Anniversary)
Salesforce, Workday, ServiceNow, SAP, IBM all shipping A2A agents. Azure AI and Bedrock both committed to A2A support.
If you’re building enterprise agents, A2A support is now table-stakes. MCP + A2A = standard stack.
|
|
5. Anthropic Hits $30B ARR in April; OpenAI Enterprise at 40% of Revenue
Anthropic: $9B (Jan 2025) → $30B ARR (April 2026). 30x in 15 months. Enterprise is where the money is.
For startups: go vertical + enterprise-first. Horizontal platforms are crowded. Vertical solutions own the margins.
|
|
Three Things That Change Starting Now
| 1. Multi-Vendor Default. Enterprises stop asking “Anthropic or OpenAI?” and start asking “MCP + A2A + which models?” Procurement changes. Architecture changes. |
| 2. Pricing Decouples from Features. The model you use stops mattering (they’re all “good enough”). The agent framework + marketplace + governance = differentiation. This is when margins compress. |
| 3. Governance Becomes Operational. A2A task auditing, permission scoping, and compliance controls move from “nice to have” to “required.” Expect SEC, SOX, GDPR to name A2A explicitly by Q4 2026. |
|
|
🔒 Premium Exclusive — A2A Implementation Scorecard
Audit Your Agent Ecosystem
The full implementation scorecard for evaluating your readiness — plus build vs buy vs lease calculator for A2A agent deployments.
| ✅ Discovery Readiness — Agent Card publication, registry integration, semantic search (0–3 months) |
| ✅ Security Posture — mTLS, scoped tokens, audit trails, compliance controls (1–2 months) |
| ✅ Protocol Alignment — MCP server coverage, A2A task management, state isolation (2–4 weeks) |
| ✅ Governance Layer — Human-in-loop, policy injection, exception handling (1–2 months) |
| ✅ Cost-Benefit Calculator — Build vs buy vs hybrid for your specific agent workloads |
$12/month. Early subscriber pricing.
Get Premium Access — $12/mo
|
|
📅 Issue #14 Preview — May 1–3, 2026
Agent SLAs: When Autonomous Fails
We’ve been shipping agents into production at scale. But what happens when an agent makes a $1M mistake? Who’s liable? Issue #14 dives into the agent-as-employee liability model, SLA enforcement across multi-vendor fleets, audit trails + provenance tracking in A2A workflows, and the new insurance products for autonomous systems (AIG, Beazley leading). Plus: case studies from three companies that hit agent failures and survived.
|
|