Cisco IQ: Engineering APIs for the Agentic Period

From day one, Cisco IQ was constructed on a transparent mandate: ship unified, deterministic, AI-powered operational intelligence by means of an API-first structure.

We delivered on that promise with the bulletins final week at Associate Summit 2026. We formally opened Cisco IQ with a complete suite of manufacturing REST APIs and our Mannequin Context Protocol (MCP) server. This ensures our intelligence runs seamlessly with any present system or AI agent.

One of the best half, the inspiration extends to our companions as nicely. Companions can now improve their very own managed companies with capabilities from Cisco IQ, built-in by means of API and MCP interfaces.

Since Cisco IQ’s normal availability in Apr 2026, over 16,000 prospects have deployed Cisco IQ, and, as we broaden this ecosystem, a brand new class of API shopper has arrived at scale: the autonomous AI agent.

Brokers don’t eat APIs the way in which human software program builders do. Builders learn specs, plan integration loops, and write deterministic code. AI brokers discover probabilistically: they guess, they retry, they usually try to hitch disparate knowledge inside their immediate context window. So, we requested a elementary engineering query: what does an API structure appear like whether it is designed to serve human builders and autonomous AI brokers equally nicely, from the very same contract?

Our reply is that trusting an agent shouldn’t be a separate choice. An agent doesn’t get a separate system, a bespoke rulebook, or an unmanaged facet door. It really works by means of the very same area mannequin and the very same safety insurance policies that already govern your engineering groups, your operational dashboards, and your enterprise integrations.

Here’s a look below the hood at how we engineered that actuality, from the area mannequin specification as much as our Mannequin Context Protocol (MCP) server.

1. One Canonical Mannequin, Each Contract

In conventional API improvement, if you wish to know which switches are affected by a selected {hardware} end-of-life announcement, a developer normally writes a customized endpoint to sew that knowledge collectively.

Our strategy begins with how infrastructure entities exist and relate to one another natively, not with an inventory of disconnected options. Cisco IQ spans a number of operational domains, and the actual worth lies within the connections between them. Once we add a brand new functionality, we mannequin the way it hyperlinks to what’s already there – bodily {hardware}, operating working programs, safety advisories, and repair contracts – and what new worth it brings to the capabilities and domains our prospects already use. Every addition makes the entire extra helpful than the sum of its components.

In apply, we mannequin the connection, not one other endpoint.

Each entity, relationship, subject definition, and enterprise rule is outlined as soon as, in a single canonical area mannequin. That mannequin is the contract. Every thing else compiles immediately from it:

• The OpenAPI 3.2 descriptions and reference documentation that builders learn.
• The underlying knowledge retailer schemas, relational tables, and indexing layers that maintain the knowledge.
• The fine-grained permission catalog that authorizes each operational name.
• The question and power contracts that AI brokers use to purpose over knowledge by means of MCP.
• Developer tooling, SDK technology, and presentation metadata that keep completely synchronized.

Relationships throughout the infrastructure lifecycle are declared explicitly within the mannequin with strict route and cardinality.

As a result of each floor derives from this shared supply, they can’t drift aside. A subject a developer inspects within the API Reference is equivalent to the sphere Cisco IQ shops within the database, indexes in question planning, and exposes to an AI agent. The OpenAPI definition acts as absolutely the boundary: if a functionality is just not within the mannequin, it doesn’t exist. And when a rule adjustments, we replace it as soon as within the canonical mannequin, and it propagates all over the place without delay: within the public REST name, in developer tooling, and within the agent’s MCP software name.

One supply of reality for folks, software program, and AI brokers.

2. Predictable for Builders, Checkable for Brokers

A well-engineered API is predictable. Study one nook of it, and you realize the remainder. As a result of each Cisco IQ API behaves the identical manner, a developer or an agent that learns one nook of it may predict the remainder, and something that goes improper comes again as one clear reply as a substitute of a guess.  We made particular architectural selections to make sure our APIs function with absolute consistency:

• Study as soon as, run anyplace: Each useful resource URL adheres to a constant construction (https://{host}/api/{namespace}/v1/{useful resource}), so no name ever must be relearned, routing by means of regional endpoints (e.g., us-west-2, eu-central-1) to implement strict knowledge sovereignty and residency mandates.
• Standardized Queries and Clear Schemas: High-level collections settle for uniform parameters utilizing the Widespread Expression Language (CEL), guaranteeing situations behave constantly all over the place. To stop SDK technology failures, we hold our OpenAPI definitions freed from oneOf schema bloat. Moreover, we implement a strict Main Underscore (_) Alias Rule for client-defined calculations in queries to ensure they by no means collide with future platform schema additions.
• Resilient Pagination and Bounded Relationships: Conventional offset pagination breaks when {hardware} telemetry streams change concurrently. Our collections implement keyset pagination utilizing opaque cursors (RFC 5988) to ensure secure reads. We additionally certain responses for to-many relationships relatively than capping knowledge, rendering huge datasets (like 10,000 safety findings) inline through bounded envelopes or as structured href hyperlinks for atomic additions and removals.
• Protected Writes and Failing Loudly: Lenient APIs encourage guessing. Unknown fields or malformed values return fast errors utilizing RFC 9457 drawback particulars, delivering all validation errors in a single structured array so a developer or agent can appropriate them in a single iteration. For updates, we use JSON Merge Patch for atomic modifications and ETags for optimistic concurrency, guaranteeing conflicting writes fail safely relatively than overwriting one other course of’s adjustments.
• Authorization within the Vocabulary of the Mannequin: Permissions are named immediately after actions, from an entire kind (learn:asset.Asset) right down to a single attribute (replace:asset.Asset.tags). Lists filter row by row, guaranteeing callers obtain solely the precise information they’re licensed to view.
• Operability and Non-Breaking Evolution: Each operation sits below an immutable model path, with responses carrying W3C traceparent headers for distributed tracing. Automated compatibility checks run in our CI/CD pipelines, catching breaking contract adjustments earlier than launch.

3. Why Wrapping APIs is Not Sufficient for Brokers

Most MCP servers at the moment expose every particular person API operation as a separate software. To reply a query like “Which crucial units have assist protection ending this quarter, and which contract covers them?”, an agent should web page by means of units, web page by means of protection information, web page by means of contracts, after which be a part of the rows itself.

In our baseline measurements, answering a pattern single cross-domain query utilizing conventional API wrapping took roughly 120 software calls simply to assemble 40 rows.

This strategy places the toughest a part of the issue – knowledge correlation throughout domains – within the place least fitted to it: the AI mannequin’s context window. Context is misplaced at each step, latency spikes, and the ensuing solutions are liable to hallucinated joins. Correlation belongs within the platform, not within the immediate.

4. Normal-Objective Operations: Shifting Correlation to the Platform

We solved this by engineering two general-purpose platform operations that deal with cross-domain complexity natively:

• Structured Question (POST /api/system/v1/question): Reads any entity with filtering, projection, ordering, and nested enlargement in a single JSON payload. It expands associated information and runs aggregations ( groupby, apply, having, compute) throughout associated domains in a single journey below the caller’s authorization.
• Bulk Engine (POST /api/system/v1/bulk): Executes a number of state adjustments in a single request. Shoppers ship a flat record of operations throughout namespaces with out computing handbook dependency graphs.

1. Automated Dependency Decision: The engine robotically orders the operations and resolves cyclic entity references by withholding deferred references and writing them as automated follow-up steps.
2. Synchronous Partial Success: Atomicity is an inherent property of the write operation, not the HTTP connection. The majority engine helps partial successes in synchronous responses, returning specific outcomes mapped precisely to the unique submitted positions. This permits a shopper or agent to immediately resubmit solely the operations that failed.

5. 4 Instruments for the Agentic Period

As a result of correlation lives within the platform, our MCP server doesn’t want to reveal lots of of fragile, particular person operations to an AI agent. It exposes precisely 4 instruments, projected immediately from the canonical mannequin:

• uncover: Finds the entities a query requires, the connection paths that join them, and the actions that apply.
• describe: Returns the precise contract, fields, knowledge sorts, and motion preconditions for an entity or operation.
• question: Validates a structured question towards the schema and executes it by means of the platform question engine.
• mutate: Executes an atomic change after validating preconditions and authorization.

When Cisco IQ provides a brand new entity, each agent discovers it robotically utilizing these similar 4 instruments. A secure, safe floor space that grows organically with the mannequin.

6. Agent Identification

An agent accessing our platform by means of MCP makes use of commonplace API authentication, inheriting the precise authorization profile of the calling principal. Cisco IQ integrates natively with Cisco Unified Identification (CUI), connecting that id to our granular, role-based catalog. Whether or not a coverage grants learn entry to a complete area or restricts write permissions solely to a single asset tag, the AI assistant operates strictly throughout the caller’s licensed boundaries. It acquires no ambient enterprise privileges just because it connects over MCP.

7. Verify Earlier than It Runs: Deterministic Conduct From Probabilistic Fashions

In an enterprise, a question that fails is an annoyance. A question that returns the improper reply is an unacceptable threat.

To make sure determinism, each question an agent writes is checked towards the area mannequin earlier than it hits the execution engine. If an agent will get a subject identify, worth, or relationship improper:

• All validation errors return without delay in a structured RFC 9457 response.
• The server features a pre-checked, corrected request payload that the agent can resubmit immediately.
• The server by no means silently rewrites or runs a question on assumptions; the agent stays in specific, deterministic management.
• Mutations current prerequisite states as learn operations first, permitting human operators to examine precise diffs earlier than confirming execution.

What does this imply in apply? In scripted exams throughout a benchmark corpus of 17 complicated enterprise eventualities, this structure delivered decisive operations enhancements:

• Whole software calls dropped from 256 right down to 36.
• Interactive approval prompts diminished from 181 right down to 1.
• Customary multi-domain learn investigations executed in precisely two calls: uncover adopted by question.

Fewer calls, fewer prompts, and solutions the platform can stand behind.

Engineering Alternatives for Prospects and Companions

These design selections ship concrete engineering alternatives throughout each operational position:

• For Prospects: Whoever acts in your behalf, whether or not an engineer, an automatic CI/CD pipeline, or an autonomous AI agent, operates throughout the precise safety guardrails you already belief. An agent by no means sees greater than the principal it represents is permitted to see, eliminating the necessity to keep parallel permission fashions. Deploying an AI agent in entrance of your knowledge is just not a brand new belief choice.
• For Companions: Combine as soon as towards a unified contract. Each functionality we introduce follows equivalent schema conventions, question parameters, and entry guidelines. Whether or not you’re constructing customized managed companies below your personal model or connecting bespoke inner brokers, you grasp the combination sample as soon as relatively than constructing one-off adapters for each lifecycle stage.

Coming Quickly: That means-Primarily based Discovery and Dwell Agent Measurements

We’re actively increasing this basis. We’re introducing meaning-based semantic discovery subsequent to key phrase search, guaranteeing pure language requests discover goal entities as reliably as queries written in strict schema phrases.

We’re additionally measuring dwell autonomous AI brokers towards our 17-question benchmark corpus, making ready to publish real-world agentic efficiency knowledge alongside our present scripted baselines.

The Basis for Agent-Prepared APIs

In the end, these architectural selections serve human builders and autonomous AI brokers concurrently. Predictable URLs, standardized response envelopes, structured error codecs, and uniform CEL parameters simplify integration for builders whereas making agent outputs strictly checkable. A single mannequin behind each contract ensures that what we expose to an agent is precisely what the platform executes.

This structure additionally prepares us for what comes subsequent. As a result of question validation, structural correlation, and error repairs dwell within the platform relatively than the immediate, execution high quality doesn’t depend upon mannequin parameter dimension.

This design permits smaller, extremely environment friendly, and self-hosted fashions to function with the identical precision as frontier fashions. It ensures that equivalent contracts and guardrails apply wherever Cisco IQ runs, from regional cloud situations to totally air-gapped sovereign digital home equipment.

APIs engineered solely for human builders are now not enough. They have to actively information AI brokers to verified truths and bodily forestall them from executing invalid operations.

That’s engineering APIs for the agentic period.

Get Began

• Dive deeper: Discover the Cisco IQ API documentation and full API Reference.
• Join an agent: Entry the Cisco IQ MCP server information to convey your AI purchasers on-line.
• Discuss to us: Contact your Cisco account group to debate the right way to leverage Cisco IQ APIs and MCP in your atmosphere.