Claude now writes its own multi-agent plans, and GitLab ships goal-driven agents
Claude now writes its own multi-agent plan, GitLab ships goal-driven agents with spend caps, and Spotify shows what doubled change volume does to verification.
Three of today’s six are really the same question: how much do you trust a loop that checks itself? Orchestration, goal-chasing agents and a faster merge pipeline all lean on it.
1 · Top story
Claude can now write its own multi-agent plan
Anthropic has added “flows” to Claude Managed Agents, in the Console. Instead of you defining a coordinator and a roster of subagents, the lead agent drafts a plan, then writes a small orchestration program that launches agents and collects their results. You can still hand it specialist subagents. This comes from a Latent Space write-up of an AI Engineer New York talk by two Anthropic engineers, not from release notes.
The engineers say flows cost much more than a single agent but find far more of the details. No pricing or GA date was given, so treat the effectiveness claim as theirs. Subagents share a filesystem and can’t message each other through the API, which keeps handoffs inspectable in the Console graph view. To test it, take one research task with a known answer set, run it as a single agent and as a flow on the same total token budget, and compare facts found per dollar.
2 · Models & features
Fine-tuned open models, judged by 26 LLM judges
Flo Health describes a post-training pipeline on Databricks: supervised fine-tuning then DPO on 32 H100s, with 15,000+ synthetic questions and 26 LLM judges (medical safety, usefulness, micro-checks), calibrated against blinded clinician review. Medical safety went from 34% to about 61%, usefulness from 42% to 76%, and the team estimates up to 10x lower inference cost than frontier APIs.
It’s a customer case study on the vendor’s own blog, so read it as a pattern, not a benchmark. The post doesn’t name the base model, report test-set sizes, or give a measured cost. The reusable part is the judge suite: if you’re weighing a small tuned model against a frontier API, build that judge set first, then compare quality and cost per thousand requests on your own traffic.
3 · Agents
GitLab’s /goal agent has a second model checking the first
GitLab 19.4 (released 17 September, covered by InfoQ on 12 October) adds a /goal command to the Duo CLI. One model makes the changes, another judges progress against the goal, and the run stops at a goal or an iteration limit. It’s public beta, on Premium and Ultimate. Expanded MCP tools let outside agents run pipelines and review merge requests, with write tools defaulting to “Always Ask”.
The InfoQ author points out that “verified” here means a second model agreed, not formal verification. The more practical bit is cost control: per-user credit caps and a per-event usage export. GitLab claims up to four times more calls per credit on hosted open-weight models (Kimi K3, MiniMax M3, GLM 5.3), with no per-model benchmarks. Pilot with a cap on one repo and compare merge-request acceptance rate against your current assistant.
4 · Research
A benchmark where coding agents sabotage each other
SWE-Duel pairs two agents on a real repository. The Red agent adds a feature and hides a subtle bug behind sealed tests; the Blue agent reviews the change without knowing a bug exists. Because tasks are generated at evaluation time, they can’t leak into training data. In an 11-participant tournament over 15 repos and five languages, GPT-5.5 with Codex CLI ranked first as a solver, repairing 98.3% of the defects it reviewed.
It’s an arXiv preprint (v1, 8 October), under review for ICLR 2027, and I couldn’t find a code release. The idea is useful anyway: an agent that writes code and one that reviews it are different skills, and your eval should probably test the reviewer on planted bugs.
5 · People worth following
Simon Willison built a file editor so secrets skip the LLM
Simon Willison released web-edit 0.1a0, a small Python web app for editing files in a folder, built with Codex and GPT-6 Astra Extra High. His reason: he wanted to edit a credentials file from his phone without sending those credentials through a model. It’s meant to run over Tailscale so an agent can start it and a human can edit.
It’s an alpha, and the post doesn’t describe auth beyond the Tailscale network. Still, the pattern is worth stealing: give agents a way to ask a human for a secret out of band, instead of pasting it into context.
6 · Architecture
Spotify’s change volume doubled, and checks didn’t keep up
InfoQ, drawing on Spotify’s engineering posts, reports monthly merged changes rose from about 8,100 to 17,000 year over year (August). Spotify’s background coding agent, Honk, runs Claude through the Agent SDK in Kubernetes pods, and Fleet Management has merged over 2.5 million automated maintenance PRs. After a 24 June content-pipeline incident and an automated dependency upgrade that passed checks but failed in production, Spotify found no material contribution from AI-written code. Volume had simply outpaced some verification controls.
The fixes are the useful part: more rollback capacity, end-to-end monitoring, and scheduling automated changes during owning teams’ working hours. Chart merged changes against revert rate for your busiest service.
Nothing here needs doing before lunch. Pick one, ignore the rest, and you’re still well informed.
Researched and drafted with AI, checked against the linked sources. How we use AI →