Pillar guide · The agent-first series · 03

What Is an Agent-First Platform?

An agent-first platform is a software platform whose relevant capabilities are designed to be discoverable, understandable and operable by AI agents within explicit permission and control boundaries. These capabilities may include reading business context, invoking workflows, changing records and tracking the outcome of an action.

In this guide, auxfirst uses “agent-first” to describe how a platform can be operated. An AI agent platform describes a broader category of tools for building and running agents. A platform can support both roles, but one does not automatically establish the other.

Format
Capability model + test
Reading time
~14 minutes
Last reviewed
September 2026
Series
03 of 04 · Platform
The short answer

A platform is agent-first when an AI agent can finish meaningful work in it — find the right capability, understand what it does, act through a programmatic interface, stay inside scoped permissions, survive failures and leave evidence of what changed — while people keep the ability to approve and stop. An API is where that starts, not where it ends.

  • Judge it per workflow. A platform can be strong for one delegated task and unusable for the next.
  • Eight capabilities, four questions. Can the agent understand, act, stay under control, and prove what happened?
  • No universal extras. MCP, long-term memory and asynchronous execution can matter; none is required everywhere.
  • Evidence over scores. Record status, evidence, gap and owner for each test instead of a single number.

01What is an agent-first platform?

An agent-first platform is a software platform whose relevant capabilities are designed to be discoverable, understandable and operable by AI agents within explicit permission and control boundaries. The definition turns on the word operable. A platform can have an excellent API and still be hard for an agent to use: operations named after a human workflow, identifiers the agent cannot resolve, side effects nobody documented, no way to tell whether a request succeeded.

Agent-first means that the capabilities that matter for delegated work were designed with an agent as one of their operators. It does not require every feature to be available to agents, and it does not remove the human interface: people still configure the platform, review what agents did and step in when needed. The claim is narrower and more useful — for the workflows you intend to delegate, an agent can do the work, and people can see and control how.

What the term is not

“Agent-first” describes a design approach. It is not a certification, a standard or a product category, and it does not mean that the human interface goes away.

This is the third layer of a series. Agent-first design decides what work to delegate, agent-first architecture describes the system that performs it, this guide covers the capabilities the agent operates, and the AI agent development lifecycle keeps the whole arrangement working as it changes.

02Agent-first platform vs related categories

Vendors use overlapping names for different things, so it helps to fix the vocabulary. This guide uses five terms as follows.

TermMeaning in this guideThe question it answers
AI agent platformTools and services for building or running agentsWhere do we build and run our agents?
Managed agent platformA service that takes over a defined part of hosting and runtime operations for agentsWhat does the provider operate for us?
Agent-first platformA platform designed so that agents can operate its capabilitiesCan an agent do real work here?
Agent frameworkA library or set of abstractions for constructing agent behaviorHow do we construct the agent's behavior?
Agent-first productA product in which an agent performs a significant part of the delegated workDoes the product do the work for its users?

↔ scroll table

The scopes overlap, and naming varies from vendor to vendor. A CRM can host its own agents, expose its records and workflows to outside agents, and ship an agent that does work for its users — all three at once. What matters for evaluation is which of those roles you are relying on.

Platforms for agents and platforms operated by agents Two overlapping areas. On the left, platforms for agents: you build and run agents on them. They include agent frameworks, AI agent platforms and managed agent platforms, and answer the question where do our agents run. On the right, platforms operated by agents: agents do work in them. They include CRMs, ERPs and service desks, developer platforms and operations tools, which are agent-first when designed for it, and answer the question can an agent finish the job here. In the overlap, both roles: a product that hosts its own agents and exposes its capabilities to outside agents. Offering one role does not establish the other. An agent-first product, where an agent does much of the delegated work, can sit on either side. TWO DIFFERENT QUESTIONS · PLATFORMS FOR AGENTS AND PLATFORMS OPERATED BY AGENTS PLATFORMS FOR AGENTS You build and run agents on them Agent frameworksAI agent platformsManaged agent platforms Where do our agents run? PLATFORMS OPERATED BY AGENTS Agents do work in them CRMs, ERPs, service desksDeveloper platformsOperations tools agent-first when designed for it Can an agent finish the job here? BOTH ROLES hosts its own agentsand exposes itscapabilities tooutside agents Offering one role does not establish the other — evaluate the role you are relying on. An agent-first product, where an agent does much of the delegated work, can sit on either side.
Figure 1 · Two different questions. A platform for agents helps you build and run them; an agent-first platform can be operated by them. Some products do both; neither role establishes the other.

If your question is where to build and run agents, the vendor landscape is mapped in Managed Agent Platforms. If it is whether agents can do work inside a platform you own or buy, the rest of this guide is for you.

03What makes a platform agent-operable?

Consider an agent working in a CRM. It needs more than a screen it can navigate. It has to identify the correct account among similar ones, understand which actions are available and what each one changes, perform an authorized operation, and confirm what actually changed. People still need ways to review, approve and stop consequential work.

Most platforms were not designed with that operator in mind. Their interfaces assume a person who can read context off a screen, recover from an ambiguous error by looking around, and remember why they clicked. An agent has none of that unless the platform provides it: stable identifiers, operations with documented effects, errors that say what happened, and a way to check the outcome.

A platform for building agents may provide the runtime for this workflow. The CRM itself also needs suitable capabilities for the agent to operate it reliably — and that is what the capability model below describes.

04The auxfirst capability model

The auxfirst Agent-First Platform Capability Model names eight capabilities and groups them under the four questions an evaluator should be able to answer for any delegated workflow. It is an analytical model developed from auxfirst's work, not an industry standard.

The auxfirst Agent-First Platform Capability Model Eight capabilities grouped under four questions around one delegated workflow, the unit of assessment. Understand, can the agent find and interpret what it needs: 01 capability discovery, which operations exist and with which parameters, and 03 structured context, the right data with its meaning. Act, can it finish the job reliably: 02 meaningful action coverage, the operations a real workflow needs, and 06 reliable execution, timeouts, safe retries and no duplicates. Control, can people bound it and stop it: 04 identity and scoped authority, attributed, limited and revocable access, and 07 human control, approve, escalate and cancel. Prove, can anyone tell what happened: 05 operational state, status, task identifiers and resumable work, and 08 evidence and inspection, results, change history and audit trail. Assessed per workflow; memory, asynchronous execution and MCP can matter, but none is a universal requirement. AUXFIRST AGENT-FIRST PLATFORM CAPABILITY MODEL · EIGHT CAPABILITIES, FOUR QUESTIONS UNDERSTANDACTCONTROLPROVE Can the agent find and interpret what it needs? Can it finish the job, reliably? Can people bound it and stop it? Can anyone tell what happened? 01030206 04070508 Capability discoveryStructured context Meaningful action coverageReliable execution Identity and scoped authorityHuman control Operational stateEvidence and inspection which operations exist, with which parametersthe right data, with its meaning the operations a real workflow needstimeouts, safe retries, no duplicates attributed, limited, revocable accessapprove, escalate, cancel status, task identifiers, resumable workresults, change history, audit trail Onedelegatedworkflow Assessed per workflow. Memory, asynchronous execution and MCP can matter — none is a universal requirement.
Figure 2 · The auxfirst Agent-First Platform Capability Model. Eight capabilities, grouped by the question each pair answers. Assessed per workflow: a platform can be strong for one delegated task and unusable for the next.
CapabilityWhat it coversExample evidence
01 · Capability discoveryHow the agent learns which operations exist and what parameters they takeA tool catalog or an API description
02 · Meaningful action coverageWhether the agent can complete a real workflow, not just readThe set of operations needed to finish the task
03 · Structured contextWhether the agent gets the right data and its meaningSchemas, identifiers, relationships, timestamps
04 · Identity and scoped authorityWhether access is attributed and limitedAccess scopes, the identity of the actor, revocation
05 · Operational stateWhether a task can be inspected and resumedStatus, task identifiers, checkpoints where needed
06 · Reliable executionHow delays and failures are handledTimeouts, retries, deduplication, events
07 · Human controlHow people approve and interrupt the workApproval, escalation, cancellation
08 · Evidence and inspectionHow execution and its effect can be verifiedOperation results, change history, audit trail

↔ scroll table

Understand01 · 03

Discovery and structured context

Can the agent find and interpret what it needs?

An agent cannot use what it cannot find or cannot interpret. A machine-readable description — an OpenAPI document, a tool catalog — tells it which operations exist; stable identifiers, typed fields and timestamps tell it what a record means. The common failure is ambiguity: two accounts with the same name, a status field with undocumented values.

Act02 · 06

Action coverage and reliable execution

Can it finish the job, reliably?

Read access is not agency. The platform has to expose the operations a workflow needs from start to finish, and they have to behave predictably under failure: clear errors, safe retries and protection against duplicate writes, for example through idempotency keys or request identifiers.

Control04 · 07

Identity, scoped authority and human control

Can people bound it and stop it?

Actions should be attributable to a specific agent acting for a specific person or role, limited by scope, and revocable without switching everything else off. Where a workflow needs it, the platform supports approval of a specific action and cancellation of work in progress.

Prove05 · 08

Operational state and evidence

Can anyone tell what happened?

Long-running work needs a status someone can check — an operation identifier to poll, or an event when it completes. Google's AIP-151 describes one well-established pattern. After the fact, someone must be able to see which calls were made, what they returned and what actually changed.

What the model does not require
  • Full parity between the interface and the API. Only the capabilities needed for the delegated tasks are assessed.
  • Long-term memory or asynchronous execution for every operation. Both matter for some workflows and not for others.
  • MCP. The Model Context Protocol can be part of the answer — its current specification includes an optional extension for long-running tasks — but exposing an MCP server does not, on its own, demonstrate complete permissions, context or reliable execution. The specification itself notes that the protocol cannot enforce its security principles; implementors have to.

05Illustrative platform examples

Three hypothetical examples — not descriptions of specific products — show how the same model reads in different contexts. Named products change quickly, so this guide stays at the level of capabilities.

Illustrative · CRM

Follow up an account

An agent reads the account's context, creates a follow-up task and checks that the task exists. It leans on structured context to pick the right account among near-duplicates, on scoped authority to create tasks without touching deal values, and on evidence to confirm the result.

Structured contextScoped authorityEvidence
Illustrative · Developer platform

Prepare a change

An agent prepares a code change, runs validation and hands the result to a reviewer. It leans on action coverage to branch, commit and trigger checks, on operational state for a pipeline that runs for minutes, and on human control — the review is the approval.

Action coverageOperational stateHuman control
Illustrative · Operations platform

Respond to an incident

An agent analyzes an incident, proposes a remediation and escalates when the change exceeds its limit. It leans on discovery of the available runbook actions, on reliable execution under partial failure, and on an escalation path that carries the evidence to the on-call engineer.

DiscoveryReliable executionHuman control

06The Agent-First Platform Test

Run the test against one specific workflow — “an agent follows up stalled opportunities in our CRM”, not “our CRM”. For each of the ten questions, record four things:

StatusSupported, Partial, Unsupported or Not applicable
EvidenceDocumentation, a test or a demonstration
GapWhat is missing
OwnerWho is responsible for closing it

There is deliberately no total score. Ten unweighted answers do not add up to a meaningful number, and a single missing capability — no way to revoke access, say — can block a workflow on its own.

Worksheet · runs in your browser, nothing is stored or sent
01DiscoverCan the agent identify the capabilities available to it?Look for: a tool catalog or an API description
02UnderstandDoes it understand the parameters, constraints and effects of each operation?Look for: documented inputs, limits and side effects
03ReadCan it access the context the work requires?Look for: schemas, identifiers, relationships, timestamps
04ActCan it perform the meaningful actions programmatically?Look for: every operation the workflow needs, end to end
05IdentifyIs it known which agent is acting, and on whose authority?Look for: an agent identity and delegated credentials
06LimitCan access be restricted and revoked?Look for: scopes, limits, revocation that does not break everything else
07ContinueCan a task continue after an interruption, if the workflow requires it?Look for: task identifiers, status, checkpoints
08InspectCan you check the calls, results and actual changes?Look for: operation results, change history, audit trail
09RecoverCan duplication be avoided and partial completion handled?Look for: idempotency, deduplication, compensating actions
10EscalateDoes the system hand the decision to a person at the right moment?Look for: approval and escalation paths with the context attached
Supported 0 Partial 0 Unsupported 0 N/A 0 Not yet assessed 10

Mark a status for each test to see the gaps here.

Figure 3 · The Agent-First Platform Test as a worksheet. It counts answers by status and lists the gaps; it deliberately produces no score. Copy the result as a Markdown table for your team.

07Adapting an existing platform

Most platforms become agent-first by adaptation rather than a rebuild. The route is incremental and starts narrow:

  1. 01

    Choose one specific workflow

    One delegated task with a clear owner and outcome. “Agent-ready” in general is not something you can test.

  2. 02

    Map the data and actions it requires

    Every read, every write and every decision point — and the identifiers that connect them.

  3. 03

    Expose the missing interfaces

    Add or document the operations the workflow needs, with their effects stated.

  4. 04

    Define permissions and approval rules

    Scopes for the agent's identity; which actions need approval, and from whom.

  5. 05

    Add statuses, identifiers and error handling

    So that work can be tracked, retried safely and resumed.

  6. 06

    Test execution and human intervention

    Run the workflow end to end, including approval, rejection, cancellation and failure.

  7. 07

    Expand the scope based on results

    Add the next workflow when the evidence from the first justifies it.

The same sequence works whether the agents are your own or your customers'. For B2B products that want agents to act inside the user's browser session, WebMCP for B2B SaaS covers one emerging route; for the product side of the change, see Agent-First Design for Legacy SaaS Retrofits.

08Build, buy or combine?

If the platform is yours, the question is how to make it operable. If you are choosing the layer that runs the agents, the question is whether to build it, buy a managed service, or combine the two. Compare the options on the same criteria, against a specific workflow — there is no universal winner.

CriterionBuildBuyCombine
Time to the first workflowLongest; everything is yours to assembleShortest, if the service fits the workflowModerate; buy the runtime, build the edges
Operational responsibilityAll of it, including runtime reliabilityShared; the provider runs a defined partSplit along a line you must document
IntegrationsWhatever you buildWhat the provider supportsThe provider's, plus your own adapters
Access and isolationFully under your controlGoverned by the provider's modelDepends on where sensitive data flows
Cost of a completed taskYour infrastructure and your peopleProvider pricing plus your integrationBoth — measure per completed task
Export of state and evidenceNativeCheck formats and completenessCheck at every boundary
Dependence on a vendorLowHigherModerate, concentrated in the bought layer
Skills to maintainRuntime, security and operationsIntegration and vendor managementBoth, in smaller measure

↔ scroll table

Whichever you choose, measure the cost of a completed task rather than a single call, and confirm you can take the state and the evidence with you. The managed options are mapped, with links to each provider's documentation, in Managed Agent Platforms.

09Frequently asked questions

What is an agent-first platform?

An agent-first platform is a software platform whose relevant capabilities are designed to be discoverable, understandable and operable by AI agents within explicit permission and control boundaries. For the workflows you intend to delegate, an agent can find the right operations, act through them, stay within scope and leave evidence, while people can approve and stop the work.

How does it differ from an AI agent platform?

An AI agent platform provides tools and services for building or running agents. An agent-first platform is one that agents can operate: its records, workflows and actions are usable by an agent within clear boundaries. One product can play both roles, but offering one does not establish the other.

Is an API enough to make a platform agent-first?

An API is the starting point. An agent also needs to discover the right operations, understand their parameters and effects, act under a scoped identity, handle failures without duplicating work, and confirm what changed. Many APIs were designed for developers who read documentation and debug by hand, and they leave those gaps open.

Does an agent-first platform need MCP?

No. The Model Context Protocol is a useful way to expose tools and context to agents and can reduce integration effort. It does not, on its own, prove that permissions are scoped, context is complete or execution is reliable. Assess those capabilities whatever the interface.

Can an existing SaaS platform become agent-first?

Yes, usually by adaptation rather than a rebuild. Start with one workflow, map the data and actions it needs, expose what is missing, define permissions and approval rules, add statuses, identifiers and error handling, test with people in the loop, then expand based on results.

Do agents replace the user interface?

No. People still configure the platform, delegate work, review outcomes, approve consequential actions and step in when something goes wrong. The interface shifts toward inspection, approval and intervention for the delegated workflows, while the rest of the product can stay as it is.

Does every workflow need persistent memory?

No. Many delegated tasks need only the context of the task and its execution state. Persistent memory helps when the work depends on history across tasks, such as account preferences, and it brings retention and correction decisions with it. Decide per workflow.

How should enterprises evaluate agent-first capabilities?

Against a specific workflow, with evidence. For each of the ten questions in the Agent-First Platform Test — discover, understand, read, act, identify, limit, continue, inspect, recover, escalate — record a status, the evidence, the gap and an owner. Avoid single scores: one missing capability can block a workflow.

10Terms used on this page

Agent-first platform
A software platform whose relevant capabilities are designed to be discoverable, understandable and operable by AI agents within explicit permission and control boundaries.
Agent-operable
Usable by an AI agent to complete delegated work: discoverable, understandable, actionable within scoped permissions, and verifiable afterwards.
AI agent platform
Tools and services for building or running agents.
Managed agent platform
A service that takes over a defined part of hosting and runtime operations for agents.
Agent framework
A library or set of abstractions for constructing agent behavior.
Agent-first product
A product in which an agent performs a significant part of the delegated work.

11Sources and further reading

Standards and specifications cited on this page

The agent-first series

Related on auxfirst

Method · This guide presents auxfirst's working model for agent-first platforms. It combines established API, identity and operations practice with our approach to delegation, authority and human supervision. The capability model and the ten-question test are analytical frameworks, not an industry standard, and the test deliberately produces no score. The three examples are illustrative, not descriptions of specific products. External sources checked on 13 September 2026.

Engage · Platform readiness

Preparing your platform for agent-operated workflows?

Work with auxfirst to identify gaps in tool access, context, permissions and operational control. For one workflow end to end, the fixed-price engagement is an AI agent readiness assessment.

Written by

Emil Krzemiński

Founder of auxfirst, the agentic experience design agency. Develops the AUX discipline and the Action Heat Ladder — the frameworks for designing trust and control into AI agents that act on people's behalf — and leads every engagement personally.

Cite this page

Krzemiński, E. (2026). What Is an Agent-First Platform? auxfirst.
https://auxfirst.com/agent-first-platform.html

Published 13 September 2026 · Markdown version

The agent-first seriesFour questions, one system.

Design defines the scope of delegation. Architecture describes the system that performs the work. Platform describes the capabilities it operates. Lifecycle describes how it is built, run and changed.