What else are people actually running?

You already have four shapes on the bench — the cloud agent on a heartbeat, the resident local agent, the same-persona federation, and the human attendant over resident units. This survey covers the other configurations companies and hobbyists are using in late 2025–2026: what each one is, who is really doing it, which design variables it sets differently, and whether a small version can run at home.

Bench rule, in your words: “I have no opinion until I've done it a few times.” One run is an anecdote, not a finding — so the shortlist below tracks repetitions, not verdicts. Evidence note: two Anthropic primary sources were read in full; everything else comes from the web index and is flagged as such in the sources.

The closest thing the field has to a settled law, 2026

Parallelise reads, not writes.

Anthropic's production Research system — an orchestrator fanning out to parallel subagents — beat its single-agent baseline by 90.2% on an internal eval, at roughly 15× the tokens of chat. Cognition (Devin), publishing a day earlier, argued the opposite default for write-heavy work: parallel agents make conflicting implicit decisions. Its April 2026 follow-up narrowed the truce — multi-agent works when writes stay single-threaded and extra agents contribute intelligence (reading, reviewing), not actions. A 2026 benchmark cited in secondary coverage (CooperBench, Stanford/SAP) found success roughly halving when a coding task was split across two agents editing shared files.

Anthropic, multi-agent research system · Cognition critique via secondary coverage — see sources.

Your starting point — the four you already know

Used here only as contrasts. Everything after this strip is an other configuration.

K1

Single cloud agent, heartbeat passes

One cloud agent with scheduled cron/heartbeat passes and tool connectors.

K2

Resident local agent

Local-LLM agent on home hardware, outbound-only, escalating to a cloud agent via shared files or its own API key.

K3

Same-persona federation

Hub-and-spoke federation of instances of the same persona, with signed tasking.

K4

Human attendant model

One human supervising several resident agent units at client sites.

The four axes that actually change an architecture

Framework names mostly re-implement the same four or five topologies with different ergonomics. These are the axes where a real difference lives — the survey's cards are organised around them.

Who owns control flow

Code, a supervisor LLM, the agents themselves, or an event stream. This decides whether behaviour is predictable, emergent, or reactive.

Where state lives

Per-agent context, a shared blackboard, a durable workflow log, or a self-edited memory store. Each brings a different failure class.

How agents reach each other

In-process calls, handoff tools, conversation, or the cross-vendor A2A protocol — from one process to strangers' agents.

What wakes the agent

A human prompt, an issue assignment, an event, a voice, or its own task list. The trigger model is often the cleanest break from K1's cron.

The bench shortlist — ordered by distance from K1–K4

Ten benches, furthest-from-known first. Effort is the report's estimate for a small home version. Tap +1 run as you complete repetitions in this session — three runs is where an anecdote starts becoming a finding.

Fold-ins, not standalone benches: C11 (MCP tool-bus) is a half-day skill to fold into benches 1, 4 and 6. C3 (code-owned workflow graph) is the assumed substrate for benches 4, 6 and 9. C4 (group chat) is worth building only as the comparison arm of bench 10, and C8 (SOP assembly line) only if document-contract pipelines are a specific interest.

    All 21 configurations

    Each card states who is actually doing it, the seven design variables — residency, topology, channel, identity & credentials, supervision, memory, trigger — and the honest home-bench verdict.

    Governance — three stacks that compose with everything above

    Not a topology, but these recur so consistently they function as configurations. OWASP's organising principle is least agency: only the autonomy, tools and access a task requires, only for as long as it requires them. Together, G1+G2+G3 on your existing benches is a half-weekend that upgrades every other configuration here.

    G1 · Approval-gate placement

    A human decision immediately before irreversible side effects — push/merge, deploy, delete, send, spend, permission change — showing the exact command and arguments, not a summary. LangGraph interrupt(), HumanLayer-style @require_approval decorators (gate in code, not in the prompt), and the rule that an agent never approves or merges its own PR. Design against approval fatigue by allowlisting low-risk actions inside a sandbox — not by rubber-stamping.

    G2 · Sandbox-per-task

    Permission prompts are not a security boundary; isolation is. Disposable container or microVM per run, no production credentials inside, no home-directory mounts, and a network egress allowlist — the main control against exfiltration after a prompt injection succeeds. Know your sandbox's scope: several products sandbox only shell commands, not file tools or MCP servers.

    G3 · Scoped, short-lived, per-agent identity

    Each agent gets its own non-human identity — service account, GitHub App, bot user — attributable and independently revocable, with scoped short-lived tokens (OIDC / secrets-manager issued, per task), zero standing privileges where feasible, and a read/write split where the platform allows. The enterprise form is Entra Agent ID (C12); the direct opposite is C18's traditional “operator's broad API key.”

    Genuinely distinct vs marketing renames

    Explicit verdicts from the survey. If you take one section into a job conversation, it's this one — most vendor taxonomies are the same menu in different vocabularies.

    Claim / nameVerdictWhy
    Supervisor vs orchestrator–worker vs CrewAI “hierarchical” vs LangGraph supervisor vs ADK root-agent delegationSAME SHAPE · C1Four vocabularies for one star topology. ADK's single-parent tree is a real constraint difference; the rest is ergonomics.
    Microsoft's five patterns (Sequential, Concurrent, Group Chat, Handoff, Magentic)CATALOGUEUseful, but not five architectures: Sequential/Concurrent = C3; Handoff = C2; Group Chat = C4; Magentic = C1 with a ledger + replanning (C5 flavour).
    Anthropic's five workflowsCATALOGUEMaps ~1:1 onto Microsoft's: chaining/routing/parallelisation = C3; orchestrator–workers = C1; evaluator-optimizer = C6.
    “Swarm” (OpenAI Swarm, Agency Swarm, LangGraph Swarm, “swarm intelligence”)ONE PATTERN, WRONG NAME · C2There is no stigmergy, no emergence, no parallelism — one active agent at a time. True swarm/stigmergy designs in production are C9 blackboards and C21 societies, rarely labelled “swarm.”
    Maker–checker vs evaluator–optimizer vs generator–verifier vs reflection loopIDENTICAL · C6Three vendors, one loop.
    Debate / council / roundtableDISTINCT PROTOCOL, BORROWED TOPOLOGY · C7C6 extended to N proposers on C4's plumbing. Only distinct if the first round is genuinely independent.
    “Ambient agent” vs Copilot Studio “autonomous agent” vs “autopilot” vs “background agent”SAME SHAPE · C15LangChain coined the useful name; later renames add nothing architectural. n8n / Home Assistant event-triggered LLM workflows are the same shape at hobbyist scale.
    “Digital employee / digital worker” (most vendors)TEST IT · C19A rename of C1/C3 inside a SaaS tenant — unless the agent holds a platform-scoped identity and works under the platform's permission/audit model. Apply that test to every claim.
    “Chief of staff” / “second brain” assistants (Lindy; Khoj)NOT NEW · ≈K1Lindy ≈ K1 (hosted single agent, connectors, inbox) with assistant-specific tooling — a valuable product, not a new architecture. “Second brain” (Khoj) is a memory/retrieval feature, not a topology.
    MetaGPT “AI software company”MOSTLY DISTINCT · C8The org metaphor is marketing, but SOP document contracts + a pub/sub artefact pool are a real design choice CrewAI-style role prompts do not make.
    Manus “general agent” / Devin “AI software engineer”PRODUCT NAMES · C13Names over C13 (sandbox-per-task) + C5 planning (+ C1 fan-out for Manus Wide Research). The VM-per-task isolation is the architecture; the job titles are not.
    MCP vs A2ANOT COMPETITORSMCP = agent→tool (C11); A2A = agent→agent (C12). Any source framing them as rivals is selling something. An A2A agent commonly uses MCP internally.
    Letta “stateful agents”DISTINCT ON ONE AXIS · C10Distinct on the memory axis only; its multi-agent features reduce to C1/C9. The contribution is self-edited tiered memory, full stop.
    Voice “AI receptionist” products (Vapi/Retell wrappers)MANAGED INSTANCES · C16(a)The cascaded-vs-speech-to-speech fork is real; the vertical branding (dental, legal, HVAC…) is not.

    Open questions — what the survey could not verify

    Carried plainly, with the hedges the research attached:

    • A2A spec version in late 2026. A stable “1.0” and its date rest on secondary explainers, not a fetched primary spec. The April 2025 launch, the Agent Card mechanism, and the June 2025 Linux Foundation donation are corroborated by multiple independent secondaries. Verify against a2a-protocol.org before building shortlist bench 1.
    • Operator / ChatGPT Agent late-2026 product status. Sources conflict (folded into ChatGPT Agent in 2025 vs later removal claims). Only the undisputed architectural fork is relied on: hosted browser vs bring-your-own-environment API.
    • Vendor performance and pricing claims — Manus GAIA scores, MetaGPT's reported 85.9% Pass@1, Agentforce/ServiceNow pricing and ARR, twin-platform pricing — are vendor- or roundup-reported and were not primary-verified. They establish that a product exists and what shape it takes, never benchmarks to act on.
    • AutoGen → Microsoft Agent Framework succession (exact 1.0 date, maintenance status) comes from secondary comparisons that agree in substance but differ in dates.
    • No hands-on verification was performed. Except for the two Anthropic posts fetched in full, framework behaviour is as documented in sources, not as tested. A bench build — shortlist #10 or #5 — is the fastest way to confirm the load-bearing claims.

    Sources

    All read 2026-11-15. FETCHED = full text retrieved; INDEX = search-index full text.