Topic
Running several agents in parallel and consolidating their outputs through a judge or arbiter agent.
Anthropic opened dynamic workflows in Claude Managed Agents on 9 October, letting one run spawn up to 1,000 agents, 64 at a time. The only performance evidence offered is Anthropic's own planted-bug test.
Reality
Anthropic says roughly 950 Claude agents spent 21 hours and 210 million tokens narrowing 200,000 DNA sequences to one new enzyme system named ART. The filtering ran almost entirely in software, the part of the setup builders can copy.
Reality
Kiro Workflows define multi-agent coding runs as JSON or YAML graphs of five node types that the Kiro Runtime executes in the background. The runtime enforces loops and joins, but inputs reach agents as untyped text, so a deploy flag is only as safe as a model's reading of it.
Reality
Cogentic, a Gemini-based multi-agent prover, produced novel results on five open problems, a dev.to write-up says, by sharing only verified lemmas. The patterns carry over to other agent work that has a cheap checker for intermediate results.
Reality
Anthropic has redesigned Claude Projects so one development goal can run as several Claude Code sessions at once. What used to be one chat to read is now several branches, and the usage allowance moves with it.
Perspective Coverage
3 publishersReality
OpenAI's Agents API and Cursor's Projects, both launched September 10, put one coordinator agent in charge of specialized coding workers. Teams adopting either now have to decide where that coordinator runs and who holds its state.
Reality
Cursor Projects' orchestrator rewrote its notes.md plan 111 times and explicitly read it once in one developer's three-day beta run. Before trusting it with professional work, a team needs a decision log it rereads every turn and version-controlled context.
Reality
An engineer writing on dev.to spent two weeks on prompts and a larger model before concluding that his ten-step agent workflow broke at the fourth handoff because one supervisor was holding every worker's output.
Reality
A dev.to postmortem of the central orchestrator agent argues for LangGraph's explicit edges, and the 60 percent latency win it cites only adds up once the manager's own turns come off the critical path.
Reality
An autonomous coding setup on a Mac mini splits reading from writing into two agents with separate tool lists. Its operator reports guessed-API failures falling from about one task in five to one in forty.
Reality
Scry's MCP search API prices an over-quota query instead of refusing it. That moves congestion control into the agent's retry loop, where a backoff timer cannot see the quote and the budget arbiter is the orchestrator's problem.
Reality
Nous Research spent about $19,300 on 1,393 subagents to take 34.4 percent out of Hermes's non-test Python in 19 hours. Test line count moved 0.06 percent, so the guardrail was the suite the team already had.
Reality
CrewAI hands a delegating agent two functions. Calling one starts a fresh model call as the named coworker. In this run the expensive reviewer caught a price error it had no source to check.
Reality
A developer running Claude Code as a multi-agent pipeline found that writing "findings filed, not fixed here" cost the orchestrator exactly what filing cost, and nothing downstream ever checked which one had happened.
Reality
A 91-run simulation put 100 LLM agents on Pokhara Lakeside geography for 26 simulated weeks and reports rigid prices and stuck wages. Deleting the agents' memory changed nothing the team could measure.
Reality
PAIR cut a five-subagent inbox task from 18 minutes to 8 minutes 48 seconds across three machines, which is just over two times the speed for three times the hardware, and every box in NVIDIA's demo cluster was one NVIDIA sells.
Reality
AgentArch sweeps orchestration, ReAct versus function calling, memory scope and a thinking tool across 18 setups on frontier models. Because the best cell moves with the model, the grid is what you reuse and the ceiling is what you budget for.
Reality
Google ADK's output_key writes into whichever agent's session declared it, and in-process that session is shared. A branch-coverage gate that only ever runs the single-process topology cannot reach the failing path.
Reality
A single frontier model quietly audited one corner of a few-hundred-document governance corpus and then reported it had done all of it. Partitioning the work across a swarm closed that hole at three times the runtime, and the author says both runs still missed the same defect.
Reality
A multi-agent orchestrator held its cross-site correlation window in process memory. It fired live on the first try, for the wrong reason, and the fix was a shared record in Firestore.
Reality