What people mean when they say "agent"
"Agent" now covers everything from a for-loop around a prompt to a governed production system. Nobody can tell them apart from the word, so here is what each vendor's own docs actually say, what the platforms actually do, and where thinqOS wins and loses. 7 August 2026.
When someone says "we built an agent," you have learned nothing. The word stretches from a few hundred lines looping an LLM through a prompt to a system that plans, acts, recovers, and is accountable for what it did. Both ship under the same noun. If you have watched the first get sold as the second, ignoring the category entirely is a rational defence.
So stop asking whether it is really an agent. Ask what state it keeps, what it can touch, who is accountable when it is wrong, and whether it can still be someone tomorrow. That is what this page compares.
Four vendors, four definitions
The most direct evidence that the word has stopped carrying information is that the four largest vendors do not agree on what it means, and each publishes its definition in its own docs.
| Vendor | What they say an agent is | What that excludes | Identity in the definition? |
|---|---|---|---|
| Anthropic | Systems "where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks" | Workflows — anything "orchestrated through predefined code paths" | No — named as an open question |
| OpenAI | "Systems that independently accomplish tasks on your behalf" | Explicitly: "simple chatbots, single-turn LLMs, or sentiment classifiers" | No — conversation key only |
| An application that "attempts to achieve a goal by observing the world and acting upon it using the tools that it has at its disposal", proactively and "independently of human intervention" | Bare models — a system without managed session history is a model, not an agent | Infrastructure layer, in preview, not in the definition | |
| Microsoft | Four different definitions across four product surfaces, from "an AI assistant that handles conversations and completes tasks" to a five-component architecture | "Traditional applications that rely on fixed rules", and separately deterministic RAG | Yes — but in the ecosystem framing, not in any of the four definitions |
| thinqOS | An identity with a Mind. Agency is a capability it can be granted, not the thing it fundamentally is | Nothing — a thinqOS identity with zero agency is still an identity | Yes — identity is the primitive |
They agree on the mechanism and disagree on the boundary. All four define an agent against the same foil, deterministic software, and two actively talk you out of using one. Anthropic: "workflows offer predictability and consistency for well-defined tasks." Microsoft, bluntly: "if you can write a function to handle the task, do that instead of using an AI agent." Credit where due.
But the boundaries conflict, and one case is worth naming. Copilot Studio's "standard harness" matches requests against authored topics with "connected steps, questions, and conditions" — and Microsoft calls that an agent. Under Anthropic's definition it is a workflow and explicitly not an agent. Under OpenAI's it is excluded outright as a chatbot. A large share of what one major vendor ships under the word does not qualify under two others' published definitions. That is not our opinion; it is three docs pages read side by side.
The thing missing from all four
Every definition above is built around the tool-use loop. None of them includes persistent identity. The industry defines an agent by what it does inside a run and says nothing about what it is between runs.
We did not catch the vendors out. Anthropic wrote it down themselves:
And on what that costs: "the core challenge of long-running agents is that they must work in discrete sessions, and each new session begins with no memory of what came before." They frame it as a harness problem to engineer around. It is a gap in the definition.
The contrast makes it starker. The interop substrate got de-vendored almost completely, and fast: MCP went to the Linux Foundation in December 2025, A2A went there too and hit a stable v1.0 in April 2026 with 150+ organisations, AGENTS.md followed, Agent Skills became an open standard. The industry standardised how agents connect to tools, talk to each other, take instruction, and package capabilities.
It standardised nothing about what an agent is. And none of those protocols carry identity — A2A explicitly connects "opaque agentic applications" sharing no memory, no tools, no internal logic. Everyone speaks the same wire protocols and means something different by the noun.
The platforms, compared
Check means first-class and documented. Tilde means partial, preview, or indirect. Dot means no. Dash means we could not verify it and refuse to guess. Arrow means shipping soon, and we only use it on ourselves — we are not going to speculate about anyone else's roadmap. thinqOS is the highlighted column; the first column stays pinned while you scroll.
| Capability | thinqOS | LangGraph | CrewAI | Google ADK | AWS AgentCore | Agentforce | Copilot Studio | ServiceNow | n8n | Zapier | Twin | Relevance | Lyzr |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Cognition — beliefs that consolidate, decay, contradict and resurface on their own | ✓ | · | · | ~ | · | ~ | · | · | · | · | ~ | · | · |
| Memory spans tools you didn't build (one mind across Claude Code and Codex) | ✓ | · | · | · | · | · | · | · | · | · | ~ | · | · |
| Durable per-user memory across sessions | ✓ | ✓ | ~ | ✓ | ✓ | ✓ | ~ | ✓ | · | · | ✓ | ✓ | ✓ |
| Belief-level confidence, decay and provenance | ✓ | · | · | · | · | ~ | · | ~ | · | · | · | · | · |
| Disclosure audience enforced at retrieval, per belief | ✓ | · | · | · | ~ | ~ | ~ | ~ | · | · | · | · | · |
| Agent designs its own database schema, under owner approval | ✓ | · | · | · | · | · | · | · | · | · | · | · | · |
| Agent's data lives in a database you own and pay for | ✓ | ✓ | ✓ | ~ | ~ | · | · | · | ✓ | · | · | · | ✓ |
| Author-your-own-graph orchestration | · | ✓ | ✓ | ✓ | ~ | ~ | ~ | ~ | ✓ | ~ | ~ | ~ | ✓ |
| Per-tool-call human approval gate | · | — | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | · | ✓ | ✓ | — |
| Cryptographic agent principal and audit trail | · | · | · | ~ | ✓ | ~ | ✓ | ~ | · | · | · | ~ | ~ |
| A2A interop (Linux Foundation spec) | → | — | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | · | · | · | ~ | ~ |
| Self-host or open source option | · | ✓ | ✓ | ~ | · | · | · | · | ~ | · | · | · | ✓ |
| Cross-vendor model portability | ✓ | ✓ | ✓ | ✓ | ✓ | ~ | ~ | ~ | ✓ | · | · | ~ | ✓ |
| Published, self-serve pricing | · | ✓ | ✓ | ✓ | ✓ | ✓ | ~ | · | ✓ | ✓ | ✓ | ~ | ✓ |
✓ first-class ~ partial, preview or indirect · no — not verified → shipping soon, thinqOS only. Rows reflect vendor documentation read on 7 August 2026.
MCP is deliberately not a row. It would be a solid line of checks telling you nothing. MCP won; it is table stakes. What differs is the shape of the support, and that is in the appendix.
The three rows that matter
Cognition is not memory. Most of this table has memory: a store you write facts into and read them back out of. Cognition is what happens to those facts when nobody is looking. In thinqOS a belief carries a confidence and a source. It strengthens with use and fades without it. Scattered specifics get consolidated into standing generalisations, which stay provisional until you confirm them. Contradictions are kept and surfaced rather than silently overwritten, and what you said directly outranks what the agent inferred. Google's Memory Bank does genuine consolidation and Agentforce scores confidence, so both get a tilde. Nobody else in that table is doing cognition; they are doing storage, and storage does not notice it is wrong.
The agent designs its own database, and cannot go rogue doing it. A thinqOS agent gets real Postgres, not a key-value bucket: it declares the shape a table should have and the platform composes the DDL, so the agent never writes raw DDL. Additive changes apply immediately. Destructive ones — dropping a column, changing a type — become a proposal its owner approves or denies. Per-subject row-level security is enforced by Postgres itself, and namespaces get limited roles that are NOSUPERUSER NOBYPASSRLS and verified after provisioning. No other platform here lets an agent evolve its own schema at all, which is why that row is a column of dots. It is also the same consent model as identity: the agent can ask to become something new, only its owner can grant it.
That database is one you own. An environment in thinqOS is an external Postgres someone else owns and pays for. AI4O's own registers through exactly the same path with no special-casing, which is the test of whether a claim like this is real. Your agent's working data does not have to live in our infrastructure.
And the honest counterweight: on durable per-user memory we have no advantage worth claiming. Google isolates memory per identity, AWS scopes it by actorId across sessions with IAM enforcement, Agentforce anchors it to a per-user profile graph, ServiceNow ships admin-defined categories. Anyone telling you those platforms are conversation-only is wrong. Our narrower claim is the row above it: their memory is scoped inside their own application — Google's bounded by app name, AWS's living inside AgentCore. A Mind sits above the tools and follows you into ones we did not build. Google takes no documented position on cross-application portability, so that is a description of their docs, not a limit they have denied.
Where thinqOS loses today
- No per-tool-call approval gate. Owners approve identity changes and destructive schema changes, but there is no general "pause before this tool runs" primitive. Google, AWS, Salesforce, Microsoft, ServiceNow, n8n, Twin and Relevance AI all ship one. This is the clearest single gap on the page and it is not on the roadmap yet.
- No self-host, no open source. Hosted only, still private preview. LangGraph, CrewAI, LlamaIndex and Google's ADK are open source. Lyzr runs fully on-premise at a published price. The agent's data can live in your Postgres; the platform cannot.
- No author-your-own-graph orchestration. Want to draw the control flow yourself? Use LangGraph or n8n. We do not do that.
- No cryptographic agent principal. Microsoft's Entra Agent ID has given every agent a first-class identity distinct from users and apps since May 2026. On the identity axis we care most about, they are ahead of us.
- No published pricing. Private preview means talking to us. Several products here take a card this afternoon.
- One arrow, not a check. A2A is a backlog item, not shipped. Until it lands, treat it as absent — that is what the arrow means, and it is the only cell on this page where we ask for any benefit of the doubt.
Platform by platform
LangGraph (LangChain)
The most control over complex, stateful, cyclical agent workflows, and the reference implementation of graph-authored orchestration. MIT-licensed and free. Its persistence model is unusually clear, splitting a per-thread checkpointer from a cross-thread Store that holds per-user facts. Worth knowing: MCP arrives through a separate adapter package rather than the core framework.
Choose LangGraph if you want to author the control flow explicitly, in code, with full open-source ownership of the runtime.
Choose thinqOS if you want memory that carries confidence, decay, provenance and disclosure scope, and that follows you across tools rather than living inside one deployment.
CrewAI
The fastest path from nothing to a working multi-agent "crew", with roles and delegation as the core abstraction, an MIT core, and a genuinely free self-host path with no execution cap. The honest caveat is production memory: the default local store has no multi-user isolation, and teams commonly bolt on an external memory provider before trusting it with real users.
Choose CrewAI if you want a role-based multi-agent prototype running today, free and self-hosted.
Choose thinqOS if per-identity isolation and disclosure control need to be properties of the platform rather than something you assemble.
Google — Gemini Enterprise Agent Platform and ADK
Formerly Vertex AI; renamed in April 2026, with Agent Engine becoming Agent Runtime. ADK is the code-first surface: Apache 2.0, multi-language, and both a client and a server for MCP and for A2A, which very few products manage on all four axes. Google also handed A2A's governance to a neutral foundation rather than keeping it. The catch is that the durable pieces — Memory Bank, the gateway, the governance suite — are Google Cloud services, so "deploy anywhere" and "durable managed memory" are not simultaneously available. Google labels its agent routing experimental, and documents that customer-managed encryption keys and the global endpoint are mutually exclusive.
Choose Google if you want one governed control plane spanning no-code business users and code-first engineers, with first-party support for both open protocols.
Choose thinqOS if the memory has to be yours and portable rather than bounded by one cloud and one application.
AWS — Bedrock AgentCore
Bedrock Agents was renamed Bedrock Agents Classic and closed to new customers on 30 July 2026; AgentCore is the successor. To be precise, because it matters: AWS calls this maintenance mode and states there is no end-of-life date and no migration deadline, so existing customers are not being evicted. AgentCore is the most protocol-neutral platform here — it hosts MCP servers, hosts A2A servers, converts your existing APIs into MCP tools, runs anyone's framework, and can switch model providers mid-session. AgentCore Policy is the strongest permission-boundary feature we found anywhere: Cedar-based rules enforced at the gateway, outside the agent's own code. AWS is also candid about its weakness, writing that "routing mode multi-agent is not straightforward today. Full multi-agent collaboration requires custom framework code."
Choose AWS if you want to bring your own framework and model, and need deterministic policy enforcement outside agent code.
Choose thinqOS if you want the cognitive layer itself rather than infrastructure to assemble one on.
Salesforce Agentforce
The strongest enterprise memory story of the incumbents: an "Agentic Memory" anchored to a persistent per-user profile graph that endures across sessions and channels, with long-term memory records carrying type, time, source, confidence and lifecycle. Native MCP client, hosted MCP servers, an MCP registry, A2A, permission-scoped retrieval and human approval gates. It is proprietary, consumption-priced, and only reachable if your business already runs on Salesforce.
Choose Agentforce if your customer data already lives in Salesforce and you want agents grounded in it with enterprise governance.
Choose thinqOS if your knowledge spans tools no single CRM owns, and you want to keep the memory when you change vendors.
Microsoft — Copilot Studio, Agent Framework and Entra Agent ID
Microsoft ships the deepest identity story in the industry and the loosest definition of the word. Entra Agent ID gives agents first-class identities distinct from users and apps, generally available since May 2026, with a registry, blueprints and governance — and it works with non-Microsoft agents. Copilot Studio brings the heaviest compliance tooling anywhere: Purview audit, Sentinel, DLP, customer-managed keys. Meanwhile AutoGen was retired into maintenance mode and merged with Semantic Kernel into Microsoft Agent Framework, which reached 1.0 in April 2026 with native MCP and first-class tool approval. The tension is the one above: some of what ships under the word here is a rules-based topic-matching bot.
Choose Microsoft if you need agents to be governed identities inside an existing Entra and Purview estate. On identity-as-infrastructure they are ahead of us.
Choose thinqOS if you want identity to mean a Mind — beliefs with confidence, provenance and audience — rather than a credential and an audit row.
ServiceNow
Now under the Otto umbrella announced in May 2026, with AI Agent Studio to build, AI Agent Orchestrator to coordinate, Action Fabric for MCP and A2A, and AI Control Tower for governance. The governance plane is the deepest we surveyed: asset inventory, approval gates, ACLs and role masking, agent lifecycle management, and risk assessment aligned to NIST and the EU AI Act. It also ships genuine long-term memory organised into admin-defined categories mapped per agent. There is no self-serve entry point and no published price.
Choose ServiceNow if you are already a ServiceNow enterprise and your compliance function has to sign off before anything ships.
Choose thinqOS if you need this for people and work that do not live inside an ITSM platform.
n8n
A visual workflow graph you can inspect step by step, with first-party MCP in both directions, human approval on individual tool calls routed to nine different channels, and a free self-hosted edition. On memory it is refreshingly blunt: "memory doesn't persist between sessions." Durable backends exist, but they are keyed by an opaque session ID, so any real user identity is plumbing you build and own. Note also that n8n is source-available under the Sustainable Use License, not open source — their words: "we do not call ourselves open source."
Choose n8n if you want to self-host, see every step, and are happy building the memory layer yourself.
Choose thinqOS if the memory layer is the part you did not want to build.
Zapier Agents
The lowest-friction way to get non-technical people shipping agents, across 9,000+ apps they already pay for, with agent-to-agent delegation and a free tier. Zapier is an MCP server. Its durable state is a curated knowledge base of documents rather than accumulated memory, we found no per-user memory documented, and no per-tool approval gate. Billing meters every activity, including each trigger and each knowledge lookup, which adds up faster than it first appears.
Choose Zapier Agents if breadth of app coverage and zero technical barrier are what matter.
Choose thinqOS if you need the agent to accumulate knowledge rather than re-read the same documents.
Twin
A Paris-based no-code platform that launched publicly in January 2026 on a $10M seed. Its differentiator is browser-first execution: agents drive real interfaces, so coverage is not gated on whether an API exists. It has the clearest per-user memory story of the newer entrants — a User Memory that "carries context across workspaces", so a new workspace already knows your preferences — and genuinely good approval gates, showing you the plan before anything runs and handing you the browser for logins and 2FA. No MCP, no A2A, no self-host, and no documented roles or admin audit log.
Choose Twin if the systems you need to automate have no usable API and your users are non-technical.
Choose thinqOS if you need open protocols, disclosure control, and an auditable record of what the agent knew.
Relevance AI
The closest comparator on this page, and the one we would hold ourselves against. Memory is explicitly scoped per user — "your memory is not shared with other users in your project" — storing role, goals, preferences and your corrections, checked at the start of every conversation. It is an MCP client and server, with documented approvals and escalations, audit logs streamed over OpenTelemetry, RBAC, SSO and data residency. Two honest limits: its agent-to-agent feature is proprietary rather than the A2A standard and supports only one-way communication today, and it is a small company.
Choose Relevance AI if you want an enterprise agent workforce with per-user memory and approvals, available today, with published governance.
Choose thinqOS if you need memory that models belief rather than profile — confidence, decay, contradiction, and who each fact may be disclosed to.
Lyzr
The deployment outlier: managed cloud, your VPC, or fully on-premise, with published per-run pricing that drops to $0.03 on your own infrastructure, and no seat licences. Its memory library, Cognis, is MIT-licensed and framework-agnostic, scoping every entry by owner, agent and session. It raised $100M in July 2026. Its A2A support is branded but we could not confirm it conforms to the Linux Foundation specification, so we have not scored it as such.
Choose Lyzr if the data cannot leave your environment and you want transparent per-run economics.
Choose thinqOS if you want the memory to be a Mind with governance built in rather than a library you wire up.
A different kind of company, and why they are not in the grid
The largest agent companies by valuation are not building platforms at all. They sell a resolved outcome — a closed support ticket, a merged pull request, a reviewed contract — and the buying decision is nothing like choosing a framework. Scoring them on self-hosting or protocol support would measure things they deliberately do not sell, so we have kept them separate rather than grading a car on wingspan.
| Company | What it sells | Scale |
|---|---|---|
| Sierra | Customer-experience agents; an Agent Data Platform unifying what a company knows about a customer across sessions and channels | $15.8B valuation, May 2026 |
| Harvey | Legal agents for contract analysis, regulatory research and litigation prep | $11B valuation, March 2026 |
| Cognition | Devin, "the first autonomous software engineer", plus its own frontier coding model | $10.2B valuation, Sept 2025 |
| Decagon | Support agents across voice, chat and email, with workflows defined in natural language | $4.5B valuation, early 2026 |
| Wonderful | Customer service agents specialised for non-English markets and local regulation | $2B valuation, March 2026 |
That the money went here rather than to the horizontal layer is itself informative. Buyers pay readily for a finished outcome in a known job. What remains genuinely contested is the layer underneath — who owns the identity, the memory and the governance that any of these outcomes depend on.
Appendix: what "native" actually means, per vendor
"Supports MCP" is now true almost everywhere and means something different in each case. Since we removed it from the grid, here is the distinction it was hiding.
| Vendor | What native means here |
|---|---|
| Anthropic | They wrote the protocol, then gave it away. MCP is a first-class connection type identically across Claude Code, the Agent SDK, the API and the desktop apps, with no wrapper and no translation step. Governance moved to the Linux Foundation in December 2025, and Anthropic states the foundation "will not dictate the technical direction of MCP." |
| OpenAI | Native means hosted and server-side. MCP is a tool type you declare in the API request, and OpenAI's own servers make the connection and list the tools. Their Connectors go further: OpenAI itself operates the MCP wrapper for third-party services. |
| Native means client-side translation. ADK's toolset class discovers a server's tools and converts MCP schemas into ADK-native formats. Note the asymmetry: consuming MCP is built in, but exposing an ADK agent as a server is documented as manual work. | |
| Microsoft | Native means governed and identity-bound. Copilot Studio syncs a server's tools and resources dynamically, but requires generative orchestration to be on and does not support prompts. In Foundry, MCP servers carry four auth modes including the agent's own managed identity. |
| AWS | Native in all three directions at once: expose your existing APIs and Lambdas as MCP tools, host MCP servers on the runtime, and consume remote MCP servers by URL with no gateway in between. |
| thinqOS | Server and client simultaneously. Outward, thinqOS bootstraps its own OAuth — walking the discovery endpoints, registering itself dynamically, using PKCE, and refreshing tokens under a lock so concurrent refreshes cannot burn a rotating token. Inward, its own agents are exposed as tools other systems can call. Every forwarded call is logged, redacted, in a ledger kept separate from model accounting. |
The bottom line
Want to draw the control flow yourself? LangGraph. Need agents inside Salesforce or ServiceNow? Use theirs, the integration is the product. Data cannot leave the building? Lyzr. Non-technical people shipping this week? Zapier or Twin. One answer does not fit, and pretending otherwise would waste your time.
thinqOS is for the case none of those cover: when the knowledge has to outlive the tool that created it, and somebody has to be accountable for what the thing knew. Every platform here has memory. Each one's memory stops at the edge of that platform, and none of them notice when what they stored is wrong. A Mind sits above the tools, carries beliefs with confidence and provenance and audience, consolidates and decays and contradicts on its own, and gives the agent a real database it can design under your approval.
We are not going to argue ours are the real agents. The useful claim is the one every vendor's definition leaves out: an agent is not just a loop that uses tools. It is something that persists, that someone answers for, and that knows what it is allowed to say. Until the definitions catch up, ask what state it keeps and who is accountable when it is wrong — and make whoever is selling you one answer both.
A point-in-time comparison by AI4Outcomes, 7 August 2026. Competitor claims are drawn from public vendor documentation read on that date; where we could not verify something, the grid says so rather than guessing.
An agent that is still someone tomorrow.
Read the point of view, or get into the private preview.