RXed AI News

What it is — in one coffee-break

Multi-agent means more than one AI agent on the same job: a lead agent splits the task, hands pieces to specialists, and stitches the results back. Three shapes turn up: handoff (support agent passes the ticket to billing), delegation (an orchestrator spawns four researchers, merges findings), debate (two argue, a third judges). The test is blunt: does the product genuinely support agents working together, or does it call one model repeatedly and print the steps with job titles on?

This sits in the Emerging row, and Emerging means unsettled. Anthropic found a lead-plus-subagents setup beat a single agent by 90.2% on their internal research eval — while burning roughly 15x the tokens of a chat, with token usage alone explaining 80% of the performance variance. Much of the "multi-agent is smarter" effect is simply "we spent 15x more".

When you actually need it (and when you don't)

Handoffs genuinely help in two places. First, read-heavy parallel work: 40 supplier price pages, 300 CVs, a dozen competitor sites. The subtasks are independent, read-only, and each fits its own context window, so nothing collides. Second, role separation where roles carry different tools and permissions — a voice agent that answers, then hands the caller to the one agent allowed to write to your calendar. That buys a permission boundary, not brains.

Theatre: anything sequential, anything with shared evolving state, and most writing and coding work. Cognition put it bluntly in Don't Build Multi-Agents: parallel subagents make implicit decisions the others cannot see, so one built a Super Mario background for the Flappy Bird clone another was drawing. The MAST taxonomy points the same way: 14 distinct failure modes across seven popular frameworks, mostly coordination failures rather than model failures.

Default for a small business: one well-instrumented agent with good tools and full logs. Add a second only when you can name the subtask, prove it is read-only, and show it runs alongside its siblings, not after.

How to recognize good vs bad implementations

Good implementations expose the seams: you see which agent ran, what it was told, what it returned, and what the orchestrator did with it. Bad ones show an animated list of role names — Researcher, Analyst, Writer — over one prompt chain in costume.

Three questions that end a sales demo quickly. Show me the trace: which agent produced this sentence, and what did it cost in tokens? No per-agent trace and no per-agent token count means you can neither debug nor budget it. What happens when subagent three returns garbage — retry, drop, or pass through? That should be a named policy, not a shrug. Can two agents run at once, or is this a queue? A queue means you pay multi-agent prices for sequential work.

One more tell: good products cap subagent count and token budget per run, and treat human escalation as a first-class handoff.

What this costs

At July 2026 rates — Claude Sonnet 5 at $2 per million input tokens and $10 output, Opus 5 at $5 and $25 — a single-agent research brief using 80k input and 10k output costs about $0.26 (EUR 0.24). The same brief as an Opus orchestrator with four Sonnet subagents, at the measured 15x multiplier, lands near $4 to $6 (EUR 3.70 to EUR 5.50). Two hundred briefs a month: EUR 48 against EUR 800 to EUR 1,100. That gap must buy something you can name.

Hosted orchestration adds a layer. CrewAI's cloud free tier allows 50 executions a month; Professional is $25 a month for 100 executions and two users, and you still bring your own LLM keys. Published per-run figures for a three-agent crew: $0.06 to $0.12 on GPT-4o-mini, $0.10 to $0.20 on GPT-4o, and $1.50 to $12 per hour in sustained production.

The self-hosted path is cheap on licence: CrewAI, LangGraph and the OpenAI Agents SDK are open source, and a VPS runs EUR 5 to EUR 20 a month. You do not escape tokens — over 90% of the bill, and self-hosting does not shrink them. Open-weight models on a EUR 1,500 to EUR 2,500 workstation delete the token line, but small models coordinate badly — the exact thing you were buying.

Where to see it scored

We audited 62 tools; 37 carry a score here. Top: Genspark (9) and Gemini / Google (9), both running genuine parallel subagents with visible per-agent output. Vapi (8) scores on voice handoffs between agents holding different tools; Claude / Anthropic (8) on subagent delegation with per-agent context isolation. The bottom is where the marketing lives: Cognism (2.5) and Base44 (2) use agent language for one model on a chain. Read Anthropic's numbers against Cognition's counter-argument: they disagree, and both are right about different tasks.