Author: Fatih E. NAR

Going Beyond Chat and Copilots

GenAI has been stuck at the MVP phase for > 2 years now.

We’ve built (and still doing) impressive chat interfaces. We’ve shipped coding copilots that autocomplete/augment our functions. AI consumer subscriptions are reaching to plateau, enterprise market seems the green grass on the other side of the fence now :-).

And the AI industry is betting 2026 will be “The year of Enterprise AI” yet most organizations still can’t answer a basic question: How do we actually plug AI agents into our business workflows?

The problem isn’t AI capability. It’s integration pattern poverty!

Previous Experiences

Enterprise architects would recognize the pain. We have solved service discovery > twice (ex; service-mesh with Istio, 5g CNF discovery with 3GPP NRF) already and failed to scale with multi-vendor inside all times.

• Service Mesh Discovery promised dynamic capability advertisement. In practice, it required rigid schema definitions and manual registration with dns service.
• 3GPP’s Network Repository Function (NRF) attempted the same for telecom network functions. Registration was static, capability definitions were controlled by a slow-moving standards body, and the system couldn’t adapt when services evolved faster than specs.

Both approaches shared a fatal flaw: Capability registration was a static, upfront act. The world moved; the registry didn’t.

Now watch the AI agent ecosystem repeat the same mistake.

What About A2A and ACP?

You might ask: “We already have Agent-to-Agent (A2A) protocol & Agent Communication Protocol (ACP). Why do we need yet another pattern?”

Here’s the difference: A2A and ACP are communication pipes. Agentic Work Exchange (AWE) is a wildlife labor marketplace.

A2A and ACP solve how agents talk to each other message formats, handshakes, point-to-point (p2p) information exchange. Essential plumbing, but they assume you’ve already decided which agents you will collaborate on what work.

That’s the gap. In enterprise reality:

• You don’t know which agent is best for a task until you test it.
• Agent capabilities evolve faster than your integration contracts.
• You want multi-vendor competition, not hardwired partnerships.

A2A/ACP are the TCP/IP of agent communication. AWE is the job board that decides who gets hired based on job needs and available budget. You need both layers but the marketplace pattern is what’s missing.

The Missing Pattern: Work Seeks Do-Ers!

What if we flipped the model?

>> Instead of registering agent capabilities and hoping workflows find them, let work find agents dynamically. Publish a task specification to a marketplace. Let qualified agents discover it, assess their fit, and compete for the contract.

This is the Agent Work Exchange (AWE) pattern.

>> AWE borrows from labor marketplaces, not service registries. When an enterprise workflow needs AI capability, it doesn’t query a static catalog. It publishes a work specification on what needs to be done, constraints, evaluation criteria, and trust requirements. Agents subscribe to work categories they are willing to make money on. Interested agents submit bids that include:

• Capability claims relevant to the task
• Proposed approach or constraints
• A reference sample a lightweight proof of competence on the actual work

The requesting workflow evaluates bids, awards a contract, and execution proceeds. No upfront registration. No schema negotiations.

AWE Pattern

Press enter or click to view image in full size

• Work Publisher: Enterprise systems emit work specifications to the exchange. Specs are semantic (natural language + structured metadata), not rigid schemas.
• AWE Bus: Message-oriented backbone. Agents subscribe to work categories. Supports both intra-enterprise (internal agents) and inter-enterprise (external agent services) with differentiated trust policies.
• Agent Pool: Heterogeneous agents from multiple vendors. Agents self-assess relevance LLMs provide the semantic flexibility that rigid capability schemas couldn’t.
• Bid Evaluator: Workflow-side logic that scores bids. Can be rule-based, LLM-judged, or hybrid. MVP samples allow evaluation on actual work, not claimed capabilities.
• Contract Engine: Manages task assignment, execution monitoring, and completion verification. Supports single-task dispatch or multi-task contract aggregation for complex workflows.
• Trust Broker: Differentiates internal vs. external agents. Manages reputation, audit trails, and compliance boundaries.

Why This Wins

>> Multi-vendor by design. No single agent vendor owns your workflows. Best-fit agents win specific tasks. Competition drives quantity & quality.

>> Enterprise-sovereign. You run your own AWE instance. The control plane stays inside your boundary. The protocol is open; the deployment is yours.

>> Scales without permission. New agents join by subscribing and bidding no registration approval, no schema updates, no integration projects. Capability evolution is continuous, not release-gated.

>> Complements, not competes. A2A/ACP handle agent-to-agent communication. MCP standardizes tool access. AWE sits above deciding which agents engage for which work. Different layer, different problem.

What’s Next

AWE is a pattern proposal, not a product announcement. The concept needs stress-testing by enterprise architects who’ve lived through the service discovery wars and see the agent integration bottleneck forming.

If you’re thinking about how AI agents will actually land in enterprise workflows not as chat toys or developer copilots, but as participants in business processes, I’d like to hear from you.