Safely scaling multi-agent teams with Claude Fable 5.2nd PostgreSQL

Anthropic researchers Barry Zhang, Tianyi Peng, Elle Najt, Keller Scholl, and Lance Martin announced that 1,000 autonomous agents running an early version of Claude Fable 5.2 reproduced a formal Lean proof of Fermat's Last Theorem. As reported in claude.dev's report, the swarm completed the formalization in 28.7 hours without researcher steering or Prove2Me scaffolding, coordinating through standard developer tooling and isolated sandboxes.

This report examines the harness architecture, shared Postgres storage, and execution boundaries used to govern autonomous agent clusters. It provides systems engineers and researchers with the benchmarks, safety constraints, and terminal commands required to run scalable coding agent teams safely.

01Formalization run with one thousand agents

The experiment tasked 1,000 agents with proving Fermat's Last Theorem, which states that no positive integers a, b, and c satisfy aⁿ + bⁿ = cⁿ for any n greater than 2. The problem stood unproven for 350 years until Sir Andrew Wiles published a 129-page proof in 1995. Anthropic supplied the agents with 60 statements to prove in Lean, consisting of the final theorem and 59 intermediate milestones.

The harness initialized the team with a one-sentence classical route sketch, mathematical definitions, and 47 proved lemmas from an earlier formalization. The setup established a hierarchy of one lead agent and 999 helpers. During the run, the harness sent the lead 19 automated reminders regarding remaining time, directing the lead to split milestones while helpers requested clarifications if assignments were unclear.

02Output token reduction and benchmark metrics

Anthropic previously formalized the proof in September using dozens of agents running an internal model comparable to Claude Fable 5.1 alongside human steering. That run required 11 days and roughly 6 billion output tokens to write 13 million lines of Lean code. By contrast, the early Claude Fable 5.2 run finished in 28.7 hours, cutting output tokens by two-thirds and producing a verified proof one-seventh the length of the original.

The efficiency gains stem from improved autonomous reasoning and persistent state tracking across the cluster. While earlier Fable 5.1 attempts failed when agents lost track of project state, Claude Fable 5.2 coordinated reliably without specialized platforms. Each Lean statement compiled independently in separate files, allowing workers to verify statements in parallel without merge conflicts.

03Shared database coordination and isolated sandboxes

The engineering team replaced custom graph orchestration with a shared PostgreSQL database and Git branches. During earlier trials, hosting 100 agents on a single virtual machine caused over half of all runs to fail due to exhausted memory, disk, or CPU capacity. To eliminate resource exhaustion, every agent received an isolated sandbox on its own lightweight virtual machine without internet access.

Agents communicated exclusively by updating PostgreSQL and opening Git merge requests into the main branch. The team selected SQL because transactional primitives handle concurrent writes from 1,000 workers and Claude understands relational queries. Database queries operated under an unprivileged role with a five-second limit and a 256 MiB write ceiling, connecting outward solely through a gateway pipe.

04Model switching commands and CLI configuration

Engineering teams deploying coding agents can install or update the command-line interface directly through the package manager. The client binary manages agent sandboxes and coordinates tool calls across active projects while maintaining environment boundaries. Upgrading ensures that the latest model definitions and safety hooks are active.

npm install -g @anthropic-ai/claude-code
claude update

Developers can switch models inside active terminal sessions or launch pinned instances directly from the command line. Default model behavior can also be established persistently across local projects using environment variables. These controls allow operators to lock agent swarms to specific model versions without altering project source code.

/model claude-fable-5-2
claude --model fable
export ANTHROPIC_MODEL=claude-fable-5-2

05Real time action monitors and kernel lockdown

Anthropic enforced a three-layer defense framework comprising hardware sandboxing, real-time tool monitoring, and offline review. Each sandbox operated in kernel lockdown mode, stripping root access pathways such as loadable kernel modules, kexec, kernel probes, and raw memory access. Sudo utilities were removed from the virtual images, preventing agent processes from modifying the outer execution loop.

A prompted language model classifier evaluated every shell command and file edit before execution. Internal evaluations demonstrated that monitors inspecting agent chain of thought and communication caught all sabotage attempts, though the production run relied on an action-based monitor to prevent persuasion risks. A violation by a helper removed only that agent, while a block on the lead halted the entire cluster for review.