Stacks/Coding agent

Replit Agent (model-directed harness)

Replit's Agent harness lets the frontier model decide effort, delegation, and which specialist to call, instead of a weaker router choosing the model. Production data shows Astra, Fable 5, and Fable 5.1 delegating in different patterns. On DeepSWE and Terminal-Bench, the shipping Max-mode agent is Pareto-efficient against Astra alone: no published Astra baseline both costs less and scores higher.

Use case

Replit's production coding agent, whose core loop decides how hard to think, when to hand work off, and which subagent (and model) should do it

Category

Coding agent

Team size

not stated

Orchestrator / framework

The main agent, or core loop, is not a separate router model: it chooses subagent tier and effort and can adjust its own effort mid-task. Harness primitives include domain-aware subagents (a general worker, read-only explorers, browser testers, reviewers, and a design subagent), small/standard/large tiers with an effort level, reusable warm subagents the core loop can return to, and dynamic mid-turn effort tuning on models that preserve the cache. The harness still decides which specialists exist; the core loop decides when and how to use them. Replit Agent never forces the core loop to spawn subagents.

Memory

Subagent context is kept so the core loop can return to a briefed subagent; the authors say nothing persists unless the model wants it back. A longer-term memory product is not described.

Evals

Evaluated in Max mode with Astra as the core loop on DeepSWE v1.1 (72% at $2.11 per task; mean of 4 reps over 113 tasks) and Terminal-Bench 4.0 (49% at $2.53 per task; 63 tasks, GPU tasks excluded). Compared with Astra alone in mini-swe-agent and with a sidekick architecture (one long-lived worker). Production delegation rates are reported for a week of traffic per model: Fable 5, Fable 5.1, and GPT-6 Astra.

Observability

A production turn from 17 September 2026 is drawn from a trace in the post (explorer, workers, tester). A named third-party observability product is not mentioned.

Monthly cost

not stated

Models

GPT-6 Astra (core loop in Max mode, and a production model)Fable 5Fable 5.1 (Claude Fable, per the authors' references)

Observability tools

not stated

Other tools

Browser-tester subagentDesign subagentmini-swe-agent (baseline only)

Source

https://frontier.replit.com/free-the-models/

Daniel Furman, Jacky Zhao, Vaibhav Kumar, Ed Sioufi, and Michele Catasta · Sep 28, 2026