Replit Agent connects to your APIs
Your agent is only as useful as the systems it can reach and the checks around its output. Today: broader API access, portable project instructions, and a practical way to trade inference compute for answer quality.
In today’s edition
- Replit Agent now connects to custom APIs.
- Claude Code adopts AGENTS.md project instructions.
- Learn self-consistency for stronger model answers.
Replit Agent can now work with your own APIs
Replit Agent now lets Pro and Enterprise teams connect APIs outside its built-in integration library, so an agent can work with the services your team already relies on.
- Add an API endpoint and authentication details, then describe how Agent should use the service.
- Custom connectors are managed at the workspace level by authorized admins and are currently in beta.
- Replit also added more than 65 audit-log events covering projects, deployments, security, connectors, secrets, and Agent activity.
- Enterprise admins can send the expanded history to the audit-log portal or configured SIEM destinations.
What it means Connect the APIs Replit does not support natively, but treat each new agent action surface as a permission and audit-log design decision.
Source: Replit
Claude Code stops making you duplicate AGENTS.md
Claude Code 2.1.277 now checks for AGENTS.md when a folder has no CLAUDE.md, giving teams a cleaner way to reuse project instructions.
- The fallback is available now in version 2.1.277.
- If Claude Code finds a CLAUDE.md, it continues to use that file instead of AGENTS.md in the default mode.
- Its Project instructions setting can use CLAUDE.md only, AGENTS.md as a fallback, both files, or managed instructions only.
- The AGENTS.md support is delivered as Claude Code’s first built-in Mod, while broader custom Mods remain upcoming.
What it means Use AGENTS.md as Claude Code’s fallback file to cut duplicate project guidance while retaining CLAUDE.md where you need it.
Source: trq212
A practical recipe for spending compute on better answers
A new hands-on inference-scaling lesson shows how to generate diverse answers, vote across them, and compare the extra compute against the accuracy gained.
- It starts by building a configurable text-generation function with temperature scaling, top-p filtering, and multinomial sampling.
- The walkthrough then adds chain-of-thought prompting, output diversity, self-consistency, and majority voting.
- It evaluates the approach on MATH-500 and covers the accuracy-versus-compute tradeoff.
- The author reports more than a 2x accuracy improvement from self-consistency and best-of-N methods.
What it means For high-value questions, sample several independent answers and aggregate them instead of betting your workflow on one generation.
Source: rasbt