AutoGPT vs Relevance AI (2026)

Stands at F and the block catalogue is far wider than the July basis conveyed. THE CLASSES ARE GENUINELY DIVERSE: an MCP TOOLS block giving access to any MCP-exposed service, a CLAUDE CODE block, an E2B SANDBOX block for code execution, image, video and music generation across Ideogram, Flux, VEO3, Fal, Replicate and MusicGen, a generic HTTP REQUEST block with multipart upload, host-scoped auth and safe-redirect protection, a read-only SQL QUERY block, Reddit blocks via PRAW, and Pinecone and Mem0. THE HTTP BLOCK'S DETAIL IS THE TELL THAT THIS IS ENGINEERED RATHER THAN LISTED. Host-scoped authentication and safe-redirect protection are things you implement after someone uses your generic request block to hit an internal endpoint they should not reach. That is a platform that has met real usage. THE MCP TOOLS BLOCK IS CREDITED HERE AND NOT ON EXT, per the standing one-line test: AutoGPT calling MCP servers is the platform consuming third-party tools. The same correction was applied to joget, wassist, agentx and aigensei this session. THE IMPORT PATH IS UNUSUAL AND WORTH CARRYING: automations import from n8n, Make.com and Zapier. That is not integration breadth as such, it is migration, and it means a customer's existing automation estate is a starting point rather than something to rebuild. Few platforms in this lane offer it. THE BLOCK DEVELOPMENT SDK REMOVES THE CEILING: customers build, register and ship their own blocks with a documented provider pattern, and AUTO-GENERATED BLOCK DOCS mean every block including custom ones is documented and searchable in the builder. A RUN AGENT block lets one agent call another as a tool, and library operation blocks add agents from the store programmatically.

Stands at F, and the multi-agent half is now documented as a mechanism rather than inferred from a visual builder. THE SUB-AGENT BLOCK IS THE LOAD-BEARING FACT: a block RUNS ANOTHER AGENT AS A SUB-AGENT WITHIN YOUR WORKFLOW, taking the agent's graph ID, version and input data, executing it and returning its outputs, with INPUT AND OUTPUT SCHEMAS DEFINING THE EXPECTED DATA STRUCTURE FOR COMMUNICATION BETWEEN THE PARENT AND CHILD AGENTS. SCHEMA-DEFINED HANDOFF IS THE PART WORTH CARRYING and it is rare in this lane. Most multi-agent claims describe agents that call each other; specifying the contract between parent and child is what makes composition reliable rather than hopeful, and pinning the version means a parent is not silently affected when a child agent is edited. The documented uses name the shapes directly: MODULAR WORKFLOWS breaking complex flows into reusable agents, SPECIALIZED AGENTS called from a main orchestration agent, and DYNAMIC ROUTING executing different agents based on input type. THE CONTROL-FLOW VOCABULARY IS COMPLETE, which few Full cells in this lane can claim: the visual builder composes blocks with branching, looping and routing logic. AI REASONING BLOCKS add an AI CONDITION block, so a branch can be decided by a model rather than only by a predicate, which is the hybrid of deterministic and reasoning control flow that outsystems and aigensei also implement. EXECUTION IS CONTINUOUS RATHER THAN REQUEST-SCOPED. Agents run continuously and on schedules, with a backend documented as carrying dedicated execution, scheduling and notification managers, so a long-running orchestration is an architectural property rather than a foreground process. What is not documented on the pages reached is failure and retry behaviour when a sub-agent errors, or parallel branch execution.

Stands at F, with the classes now named rather than summarised as continuous operation. THREE CLASSES ARE DOCUMENTED AS BLOCKS. WEBHOOKS give the inbound event class, so an external system fires an agent directly. SCHEDULING blocks give the recurrence class, and the backend is documented as carrying a dedicated scheduling manager rather than a cron field bolted on. The public REST API gives the inbound programmatic class, with agent execution exposed as an endpoint. CONTINUOUS OPERATION IS A FOURTH SHAPE and it is the platform's distinguishing one: agents RUN CONTINUOUSLY rather than being invoked and torn down, with the backend documented as including execution, scheduling and notification managers and websockets for real-time communication. A continuously running agent is not waiting for a trigger in the usual sense; it is already running and reacting. THE LIBRARY BLOCKS ADD AN INTERNAL ROUTE, since a sub-agent block lets one agent start another, so work reaches an agent from another agent as well as from outside. WHAT IS ABSENT is a conversational channel layer: no chat, email, Slack or messaging surface through which a person addresses an agent is documented, beyond the platform's own interface. Reddit blocks are an integration target rather than a channel a person uses to reach an agent. For a developer-facing builder that is coherent, and the event, schedule and API classes carry the grade comfortably without it. Notification managers are documented in the backend, which supports the HITL notification path credited on that axis rather than counted here.

P>F. The July basis said data and knowledge blocks exist within workflows but no first-class RAG system is documented. A named vector-storage block category is that system. PINECONE VECTOR STORAGE IS A DOCUMENTED BLOCK, listed under a VECTOR AND MEMORY category alongside Mem0. A vector store block means the customer builds and maintains a persistent index the agent queries, which is a maintained retrieval structure rather than context assembled per request, and it clears the persistence line the same way joget's Pinecone Tool did earlier today. Grading these two differently would be inconsistent. THE INGESTION AND TRANSFORMATION SIDE IS UNUSUALLY WIDE, which is what turns a vector store into usable grounding: a spreadsheet reader, XML parser, text encode and replace, and list and dictionary manipulation, plus an HTTP Request block with multipart upload for pulling documents in, and a read-only SQL Query block over the customer's own platform data. MEM0'S KNOWLEDGE BUILDING adds a second grounding route, accumulating facts and insights over time into a growing base the agent draws on. That is credited primarily on Mem as accumulation, and noted here because the two axes genuinely overlap in this architecture; the vector index is the grounding structure and the memory layer is the state, and I have kept the Pinecone fact here and the Mem0 fact there. AI REASONING BLOCKS INCLUDE A FACT CHECKER, which is a grounding-adjacent quality control worth recording. CONFIDENCE IS MEDIUM AND THE LIMIT IS THE SAME AS ON JOGET: Pinecone is a customer-supplied dependency rather than a bundled knowledge base, and no chunking strategy, embedding model, retrieval configuration or citation behaviour is documented. What the platform ships is the connection and the composition surface, not an opinionated RAG pipeline. For a developer platform that is the coherent trade.

P>F, and it clears both limbs of the 31 August ruling with named blocks rather than claims. MEM0 IS A FIRST-CLASS BLOCK FAMILY, and the vendor's own block documentation states the property directly: MEMORIES PERSIST ACROSS WORKFLOW EXECUTIONS, ENABLING YOUR AGENTS TO REMEMBER PAST INTERACTIONS. Across executions is the ruled Full condition stated in the vendor's words. The documented use cases name the shape precisely: CONTEXT CARRYOVER, saving details from one run to reference in subsequent runs without re-asking; PERSONALIZED ASSISTANTS, storing user preferences and learned information for future conversations; and KNOWLEDGE BUILDING, ACCUMULATE FACTS AND INSIGHTS OVER TIME, which is the accumulation limb explicitly. THE BLOCK SET IS OPERATIONAL RATHER THAN A SINGLE STORE CALL. Separate blocks add memories with metadata for filtering, retrieve all memories scoped to a run or an agent, and fetch the latest memory with key filtering. The documentation names debugging WHAT YOUR AGENT HAS LEARNED and identifying OUTDATED OR INCORRECT INFORMATION THAT NEEDS UPDATING as use cases, which means the memory is inspectable and correctable, the property missing from every other Full memory cell reviewed this session. A SECOND, SIMPLER PRIMITIVE SITS ALONGSIDE: a KEY-VALUE STORAGE BLOCK described as a persistent KV store ACCESSIBLE FROM ANY AGENT TO SHARE DATA ACROSS TASKS AND AGENTS. Cross-agent shared state is a different capability from per-user memory and both are documented. UNDER THE RULING, customer-wired persistence from documented primitives counts, and this is the clearest instance of that in the lane: Mem0 is a purpose-built memory layer, exposed as blocks, with retention scoping and inspection. What is not documented on the pages reached is a retention period or whether Mem0 is bundled or requires the customer's own account.

Stands at P, and the July basis was right that no formal approval gate framework is documented. WHAT EXISTS IS REAL BUT PARTIAL. Notifications fire when an agent needs input, and the backend carries a dedicated notification manager, so a person can be pulled into a run. Run pause and edit controls let an operator stop and change a running agent. A user input block accepts and validates input values within a workflow, so a workflow can be authored to wait for a person. THAT LAST ONE IS THE STRONGEST AND IS ALSO THE LIMIT. A customer can build an approval step by placing an input block before a consequential action, which means approvals are achievable. But this axis asks what the platform ships, not what a determined builder can assemble, and the same reasoning kept joget's Pack at Partial and refused BeanShell extensibility on its Comp cell. A generic input block is a primitive, not an approval mechanism: nothing routes to a named approver, records a decision, times out, or distinguishes an approval from any other input. THE GUARDRAIL HALF IS THINNER STILL. Host-scoped authentication and safe-redirect protection on the HTTP block are real constraints on what an agent can reach, and a Fact Checker block exists among the AI reasoning set. No content policy, spend limit per agent, action allowlist or confidence threshold is documented. THE COST CONTROLS ARE WORTH RECORDING as an adjacent form of restraint: per-agent, per-run and per-day spend visibility with cost per block, on a credit wallet. That bounds runaway behaviour commercially rather than by policy, which for continuously running agents is not nothing. CONFIDENCE MEDIUM: the block catalogue is large and I read categories rather than the full list, so a dedicated human-in-the-loop block could exist. That is the check at lane close, alongside the Stagehand block list.

Stands at P, and it is the one cell where this record's open-source, developer-first character costs it rather than helps. WHAT IS DOCUMENTED IS ENGINEERING PRACTICE, NOT POSTURE. Credential and integration management APIs handle secrets, CORS origin allowlisting bounds who can call the platform, and the HTTP Request block carries HOST-SCOPED AUTH AND SAFE-REDIRECT PROTECTION. That last pair is a real control against an agent being used to reach internal endpoints, and it is the kind of thing implemented after someone tries it. Those are genuine, and they are why this is Partial rather than None. WHAT IS ABSENT IS THE WHOLE OF THE CONJUNCTION'S FIRST HALF. No SOC 2, no ISO, no trust page, no attestation, no report path, and no enterprise identity: no SSO, no SAML, no SCIM and no role-based access control appears on any page reached. For a platform selling Team subscriptions that is a notable gap, since a team plan without RBAC means every member has the same reach. THE SELF-HOSTED PATH CHANGES THE QUESTION RATHER THAN ANSWERING IT, and this is worth stating for a reader. A customer running the platform on their own infrastructure inherits their own controls: their identity provider fronts it, their network policy bounds it, their audit covers the host. That is a real answer for a security-conscious buyer and it is graded on Dep where it belongs. It does not make the platform's own control surface any deeper, and a hosted-cloud customer gets none of it. THE HONEST READ FOR A BUYER: this is a capable developer platform with an enterprise-readiness gap that is consistent across Sec, HITL and Eval, the three cells at Partial. Those are precisely the governance axes, and a 12.5 composite with exactly those three soft is a coherent and informative profile rather than an anomaly.

Stands at F and is one of the strongest Obs cells in the lane, with a property no other record reviewed this session has. THE SQL QUERY BLOCK IS THE DISTINCTIVE FACT: READ-ONLY SQL ACCESS TO YOUR PLATFORM DATA, INCLUDING CUSTOM ANALYTICS VIEWS. Every other Full Obs cell in this batch rests on a dashboard the vendor built and a claim about what it shows. Here the customer can query the execution record directly and build their own views over it. That answers questions the vendor did not anticipate, which is the difference between reporting someone designed and auditability you own. It is also composable: the SQL block sits in the builder, so an agent can query the platform's own run history as a step. COST GRANULARITY IS THE SECOND UNUSUAL PROPERTY. The dashboard reports dollar cost PER TASK, NODE, AND BLOCK, alongside per-agent, per-run and per-day spend. Attributing spend to an individual block inside a run is finer than most platforms attempt, and for a pay-as-you-go credit model it is the number a customer actually needs. PER-RUN INSPECTION covers the reconstruction half: the dashboard is described as mission control with per-run inspection and execution analytics, and diagnostics dashboards are documented. AUTO-GENERATED BLOCK DOCS support the explainability side indirectly, since every block including custom ones carries documentation surfaced in builder search, so a reader of a run can find out what each step was. What is not documented on the pages reached is a retention period for run history, an export path, or whether a reasoning trace exists for AI blocks as distinct from the input and output record. The block-level structure means the path through a graph is reconstructible even without one.

Stands at F, and like joget this is the strong form of the axis because self-hosting is a right rather than a tier. THE PLATFORM IS OPEN SOURCE AND SELF-HOSTS FREE via Docker Compose, with the customer supplying their own model API keys. No sales conversation, no enterprise plan and no gate: a customer clones the repository and runs it. Two records this session were held at Partial for asserting on-premises deployment with no route a customer could take, and this is the opposite case. THE RESIDENCY LOOP CLOSES COMPLETELY, which is the test that separated the strong Dep cells from the weak ones across this batch. Because the customer supplies model keys on the self-hosted path, inference calls go from their infrastructure to a provider of their choosing, and with Llama in the pre-integrated set the model can run locally too. Neither the platform nor the inference need touch vendor infrastructure. Most platforms claiming self-hosting still route model calls through their own service, which quietly defeats the point. THE HOSTED CLOUD IS THE ALTERNATIVE PATH with Pro, Max and Team subscriptions and a credit wallet, so the same product is available managed for customers who do not want to run it. ONE LICENSING CONSTRAINT RECORDED HONESTLY: platform code is under Polyform Shield while the wider repository is MIT. Polyform Shield permits use, modification and self-hosting but bars building a competing product, so a customer's freedom to run and change it is real while the licence is not OSI-open in the way the repository's headline suggests. That distinction matters to a buyer doing licence review and should not be smoothed over. What is not documented is region selection within the hosted cloud, or any data residency commitment for cloud customers.

Stands at F and is the clearest instance of the 31 August Pack bar in the lane, because the catalogue is not only browsable but programmatically addressable. THE MARKETPLACE IS A REAL STORE, not a template gallery: community-built agents with RATINGS AND REAL RUN COUNTS, plus one-click adds to a personal library. Run counts are the property worth carrying. A rating tells you what people thought; a run count tells you what people actually used, and it is the number that separates a marketplace with traffic from a shelf of published intentions. Very few catalogues in this index expose it. THE LIBRARY OPERATIONS BLOCKS TAKE THIS FURTHER THAN ANY OTHER RECORD REVIEWED. A block ADDS AN AGENT FROM THE STORE TO YOUR PERSONAL LIBRARY using its store listing version ID and returns the library entry and agent graph IDs, which can then be used to execute it. Another retrieves all agents in the library with search and pagination. Adoption of a packaged agent is itself an automatable step, and the documented uses are AGENT PROVISIONING, automatically adding recommended agents to a user's library, ONBOARDING FLOWS seeding starter agents, and DYNAMIC AGENT ACCESS adding agents on demand when a user requests a capability. THAT LAST ONE IS ARCHITECTURALLY NOTABLE: an agent that acquires new agents at runtime in response to what a user asks for. The pack layer is composable with the runtime rather than sitting beside it. VERSION PINNING IS PRESENT via store listing version IDs, so an adopted agent is a fixed artefact rather than a moving target. The n8n, Make.com and Zapier import path adds a further route to working automations and is credited on Int.

Stands at F and is close to the strongest form this axis takes, because model choice is per block rather than per account and the customer holds the keys. PRE-INTEGRATED PROVIDERS ARE OPENAI, ANTHROPIC, GROQ AND LLAMA, with third-party integration guides showing further providers connected through AI blocks. Groq and Llama together are the entry that matters: Groq is a low-latency inference route and Llama is open weights, so the set spans hosted frontier models, fast inference and models a customer can run themselves. PER-BLOCK SELECTION IS THE ARCHITECTURAL POINT. In a graph where each node is a block, choosing the model at the block means a single agent can run a cheap fast model for classification and an expensive one for reasoning in the same workflow. Platforms that set one model per workspace cannot express that, and here it is the natural consequence of the block design rather than a feature someone added. BRING YOUR OWN KEYS IS THE COMMERCIAL SHAPE and it is unqualified on the self-hosted path: self-hosting is free and the customer supplies their own model API keys, so the platform is genuinely indifferent to the provider and takes no margin on inference. On the hosted cloud a credit wallet covers agent runs. THE CONTRAST WITHIN THIS SESSION IS INSTRUCTIVE: play-fast sits at None on a single named provider because its premise is a closed governed environment; AutoGPT sits at Full because its premise is that the customer assembles everything. Both grades describe a product decision rather than a deficiency. No automatic runtime routing between models is documented, which this axis does not require since it measures customer choice.

Stands at F and clears Mike's 30 August bar several times over; it is among the most complete extensibility surfaces in the lane. THE PUBLIC API IS DOCUMENTED PROPERLY: a REST API with an OPENAPI SCHEMA and GENERATED SDK CLIENTS, supporting agent creation and execution programmatically. A published schema with generated clients is a materially stronger claim than an API mentioned in a feature list, which is where several records this session sat, because it means the surface is machine-described and the clients are derived from it rather than hand-maintained. THE BLOCK DEVELOPMENT SDK IS THE SECOND SURFACE AND IT IS THE MORE INTERESTING ONE. Customers BUILD, REGISTER, AND SHIP THEIR OWN CUSTOM BLOCKS using a documented provider pattern, with a ProviderBuilder configuring API keys and per-run costs. Extending the platform's vocabulary rather than only calling it is a deeper form of extensibility, and the auto-generated documentation means a custom block is a first-class citizen in builder search alongside shipped ones. THE OPEN-SOURCE CORE REMOVES THE CEILING ENTIRELY. The repository is public, self-hosting is free via Docker Compose, and a customer can read and modify the platform. Platform code is under Polyform Shield with the wider repository MIT, which is a real licensing constraint worth recording rather than eliding: Polyform Shield permits use and modification but bars building a competing product. WEBHOOK BLOCKS provide the inbound event surface and are credited on Trig. THE MCP TOOLS BLOCK IS DELIBERATELY NOT COUNTED HERE. AutoGPT calling MCP servers is the platform consuming tools, which is Int under the one-line test. No first-party MCP server exposing AutoGPT agents to an outside assistant was found, which under the ruling does not withhold the grade given the REST API and SDKs.

Stands at P, and the shape of this cell is unusual: the platform is strong on the debugging limb and absent on the testing one. DEBUGGING AND OPTIMISATION ARE WELL SERVED. Per-run inspection, execution analytics and diagnostics dashboards let a builder see what a run did, and cost reported per task, node and block makes optimisation quantitative rather than a guess. The SQL Query block credited on Obs extends this: a builder can query their own execution data and build custom analytics views over it, which is a genuine optimisation instrument. The block structure itself aids debugging, since a failing step is identifiable by node rather than inferred from a final output. THE TESTING LIMB IS WHERE IT STOPS. agbenchmark exists in the repository for benchmarking agents, but it belongs to the original autonomous-agent project rather than to the current block platform, and nothing documents a test set, expected outputs, a scored run, a judge or a comparison between agent versions on the same inputs. The ruled bar asks for a result the customer can read AND COMPARE about agent behaviour on their own work; per-run analytics give reading without comparison. THE GAP IS SHARPER HERE THAN THE GRADE IMPLIES, and it is worth carrying. This platform runs agents continuously, lets model choice change per block, imports workflows from three other tools and adds community agents from a marketplace at runtime. That is a lot of surface across which behaviour can shift, and no regression check is documented for any of it. A marketplace agent adopted by version ID is pinned; an agent whose model or blocks a builder edits is not. CONFIDENCE MEDIUM: the block catalogue is large and read by category, and a testing or assertion block could plausibly exist among the AI reasoning set, where a Fact Checker already does.

N>F. The July basis said no first-class browser or computer use was documented ON THE CURRENT PLATFORM, which was the right question to ask given AutoGPT's history, and the answer has changed. THE DOCS CARRY A STAGEHAND BLOCKS SECTION. It sits in the block-integrations tree, immediately before System Library Operations in the documentation navigation. Stagehand is Browserbase's AI browser automation framework, built on three primitives, ACT to perform an action described in natural language, EXTRACT to pull structured data from a page, and OBSERVE to identify what is actionable. It runs on top of Playwright with an LLM resolving intent against the live page. ACT IS THE PRIMITIVE THAT DECIDES THIS. Clicking, typing and selecting on a live page is operating software through its interface, not retrieving from it. Extraction alone would be scraping and I have refused it repeatedly this session; act plus extract sits on the other side of that line, and it is the same distinction that carried kalcend and integrail to Full. SHIPPING IT AS A BLOCK CATEGORY IS WHAT MAKES IT FIRST-CLASS. Browser control is not a bolt-on here; it is a set of composable blocks that drop into the same visual builder as HTTP requests and database queries, so a workflow can call an API where one exists and drive a page where one does not. That is exactly the case this axis is for. CONFIDENCE IS MEDIUM AND THE LIMIT IS PRECISE: I established the section from the documentation navigation of a first-party docs page rather than by reading the block list itself, so which Stagehand primitives are exposed as blocks, and whether a hosted browser is included or the customer supplies Browserbase credentials, is not established. Reading that page is the check at lane close. An E2B sandbox block also exists; per the standing convention code execution in a sandbox is not computer use and is credited on Int.