Lire cette page en français

Agents

From tasks to outcomes: Devin Bots, Replit Missions and Google’s Stitch Loop; GitHub CLI learns to wait for PR events; Claude Code 2.1.328

Cognition, Replit and Google Labs all launched agents that own a goal instead of a single task: Devin Bots lead teams of Devins, Replit Missions run agent swarms under a hard budget, Stitch Loop watches your metrics and opens PRs. GitHub CLI 2.105.0 adds gh event listen so agents stop polling, Copilot gets eight new MCP connectors, and Claude Code 2.1.328 ships.

By ·

Between November 17 and 18, 2026, three vendors launched the same idea under different names: you hand an agent an outcome, not a task, and it plans, delegates and checks the work until the goal is met. Here is what each one actually ships, and the figures are the vendors’ own.

Devin Bots: an agent that manages Devins

Cognition introduced Devin Bots on November 17 (Cognition). A Bot is a persistent agent with its own memory and tools that starts and monitors several Devin sessions in parallel, each on its own machine. According to the documentation:

Bots come with the Max plan at $200 a month and with team plans (Devin pricing). Cognition says an “Android Bot” built Devin’s Android app from 1,217 tickets it filed itself, and that a “Layout Bot” got 159 pull requests merged in three weeks, cutting lag by 70%. Fortune reports that a Bot can join a team’s Slack and work over iMessage, and puts Cognition’s valuation at $48 billion (Fortune).

Replit Missions: an agent swarm with a spending cap

Replit launched Missions on November 17 (Replit). You agree up front on the goal, the success criteria, a maximum budget and the actions that need your approval; a swarm of agents then plans, builds and tests, and an independent AI judge checks the result against your criteria. Any increase in the limit needs your sign-off, and messages you send during a mission count toward its budget (Missions page).

Replit’s figures: a swarm built DataCat, an in-house monitoring platform, in two days on a $3,000 budget. On 13 NumPy optimization tasks in Auto (Max) mode, one agent reached speedups of up to 4×, a 16-agent swarm 17× and a 32-agent swarm 19×. Replit has not published a separate price for Missions or the plans that include it.

Stitch Loop: a product agent that watches your metrics

Google Labs announced Stitch Loop on November 18 as a new direction for Stitch (Google). You connect GitHub and your product context (logs, user behavior, feedback) through more than 180 integrations, custom MCP servers or CLI uploads. Three loops come by default (cybersecurity, latency, accessibility) and you can add your own. Each loop suggests, proposes solutions, writes the code in an isolated cloud sandbox and opens a GitHub pull request; nothing merges until you review it. Stitch Loop is behind a waitlist.

GitHub: gh event listen and eight new Copilot connectors

GitHub CLI v2.105.0 adds gh event listen: give it a pull request URL and the events you want (submitted review, state or head change, merge conflict, completed check suite) and it prints each event as a JSON line instead of making your agent poll (GitHub). It exits after the first event by default; --count, --follow and --timeout change that. The gh event commands are in preview and need access to Event Streams, which is on a waitlist.

Copilot also gets eight GitHub-verified MCP connectors in public preview on all plans: Box, Buildkite, Calendly, ClickUp, Mixpanel, Supabase, Webflow and Wix, on top of more than 25 existing ones (GitHub). Team customizations are now generally available on Copilot Business and Enterprise: team maintainers can turn on approved Agent Plugins by default for their team (GitHub).

Claude Code 2.1.328

Published late on November 18 (UTC), Claude Code 2.1.328 brings auto mode to Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud’s Agent Platform, Microsoft Foundry and the Claude apps gateway, adds "onFailure": "block" for mcp_tool hooks, and adds self-hosted runner options (--repository-folder, --hook-label, --metadata). Among the fixes: zsh commands stopping at a word that starts with =, automatic compaction reusing the old system prompt, and several Code Mode issues on resumed sessions (release notes).

What we updated

Sources

Spotted an error in this article? Report it · Our corrections policy

OrigoTech logo

New: the app iPhone & iPad

OrigoTech is now an iPhone and iPad app

Read the articles natively, with “New” markers on unread ones, local search and offline reading of what you have already loaded. Free, no account, iOS 17 or later; ads appear between articles.

Rekha icon: a house drawn in a single stroke

From the same author Self-promotion

An Eulerian path, drawn with your finger

Rekha, the iPhone game by OrigoTech's author: connect every dot in a single stroke without tracing the same link twice, across 500 drawings that are all solvable.

The OrigoTech newsletter

A sourced weekly digest of AI tool news.

Read next

Agents

GitHub at Ignite: Copilot Evals, “Adaptive” code review and Copilot in the browser; Foundry adds Baseten; TensorWave commits 457 MW to OpenAI

At Ignite, GitHub puts Copilot Evals in public preview to test your skills, agents and MCP tools, lets Copilot pick the effort for each code review and lets it drive Chrome. Microsoft Foundry adds Baseten and deploys models and agents onto your own AKS clusters. TensorWave will supply 457 MW of AMD capacity to OpenAI. Also: Claude Code 2.1.327, Codex CLI 0.170.0, OpenCode 2.3.2.

· 7 min read

Agents

Eclipse Enclave sandboxes coding agents, OVHcloud launches 55 AI products with B300 and Rubin GPUs, OpenAI shows its Codex-run Performance Factory; Claude Code 2.1.329 and Codex 0.171.0

The Eclipse Foundation releases Enclave 1.0, an MIT-licensed sandbox that runs Claude Code, Codex or OpenCode in isolated containers; Kong launches Volcano; the GitHub Copilot app reaches Azure DevOps repos. OpenAI says Codex agents ran more than 1,700 performance investigations in about three months, and ships GPT-Live-1.1 in its API. OVHcloud unveils 55 products and up to 60,000 B300 GPUs, Tencent publishes the Hy4 weights, and Claude Code 2.1.329, Codex 0.171.0 and Zed 1.29 ship.

· 6 min read

Agents

JetBrains runs Codex and Claude Code inside TeamCity builds; Hugging Face hands its inference routing to OpenRouter; ngrok and Semgrep ship MCP servers

JetBrains released an experimental TeamCity plugin that installs Codex or Claude Code on the build agent, lends it credentials only for the build and records tokens and trajectories. Hugging Face now routes chat requests through OpenRouter, with failover and no markup. Also: ngrok's MCP server, Semgrep's agentic platform, Claude Code 2.1.326 and OpenCode 2.3.0.

· 5 min read