Software delivery teams

Four-agent operating team: orchestration, engineering, operations, independent audit.

Claude, Hermes and Codex agents that research, build, review and ship agentcv.ai. One human owner approves every release.

Planner and coding agents built into Atlassian's internal Jira that take a work item to a pull request, with an engineer approving each step.

GitHub Next's fleet of more than 100 narrowly scoped agentic workflows that look after the open-source gh-aw repository, from issue triage to CI diagnosis.

Spotify's internal agent setup for code changes: an interactive agent scopes the task, a background coding agent opens the pull request, and an LLM judge checks the diff.

Specialist agents analyze team designs before an orchestrator reconciles their findings for engineer review.

Claude Code roles take tasks through specification, implementation, testing and final acceptance.

Seven Claude Code subagents in separate context windows, including product, backend, launch and a hostile QA reviewer, used to plan a solo product.

A developer-creator's everyday agent setup: Codex App for all development, Claude Code for UI work and a Telegram assistant, plus scheduled and overnight agents.

A review coordinator assigns merge-request checks to specialists in quality, security, compliance, documentation, performance and release impact.

Internal reviewer and judge agent pair that comments on HubSpot pull requests and drops low-value feedback before anything is posted.

An engineer's personal agent fleet: a Council workflow plans each ticket, Andúril runs mostly Claude-based implementers, and Codex and blank-context agents review the result.

Two OpenClaw agents sharing one Oracle Cloud server: a memory-backed assistant for newsletter inbox work and a website-builder bot used with a business partner.

Project-specific coding specialists coordinated through a main Claude Code session.

Independent agents the Next.js maintainers use to research, reproduce, verify and fix GitHub issues; they helped cut the backlog by 1,462 closed issues in about three weeks.

Specialized Claude Code review agents check pull requests before human reviewers assess design.

A product-manager agent and a coder agent, built from the author's own Bobbo agent code, that take turns running his MCP deals server.

Internal agent toolkit Shopify uses to port its Shopify app from React Native to native, with an orchestrator, an implementer, a test-writing subagent and independent reviewers.

Claude Code builds features from specs and opens pull requests, Codex reviews and tests them, and the Asian Efficiency founder approves the merges.

Coordinator + reused sub-agents across two developer-platform tools.

Agents coordinating repository changes for software evolution.

Open-source AI software developer with sandboxed runtime — ICLR 2025.

Non-agentic 3-stage pipeline — 40.7% Lite / 50.8% Verified with Claude 3.5 Sonnet.

Two-role pipeline: reasoning architect + format-specialist editor — 85% on Aider's code-editing benchmark.

Central LLM dynamically breaks down tasks and delegates to specialist workers.

Flexible multi-agent group conversation — hierarchical, peer, or proxy topologies.

5-role sequential pipeline — 22,949 tokens, 148s per software task.

Swarm teammates sharing a task list — parallel exploration on a single codebase.

Lead agent spawns specialist subagents in isolated context windows.

Single-agent software engineer achieving 49% on SWE-bench Verified.

5-agent supervised pipeline — 28.33% SWE-bench Lite, $3.09/issue.

Vendor-reported #1 on Terminal-Bench Core v0.1.1 at publication (Sep 2025, 58.75%) — model-agnostic harness.

Microsoft Research generalist 5-agent system: GAIA 32.33%, WebArena 32.8%.

Multi-objective Bayesian search for team config — >45% cost reduction vs random.

39.33% on SWE-bench Lite (lab-page claim).

5-agent SOP-encoded pipeline — 124.3 tokens/LoC, executability 3.75/4.

Minimal bash-only agent — reference floor on SWE-bench Verified.

Solo agent, multi-model orchestration — 70.4% on SWE-bench Verified.

Python-first multi-provider agent — minimal 1K-line library, MCP + LangChain tools.

Solo software agent with custom ACI — 12.5% SWE-bench, 87.7% HumanEvalFix.

Three-lens parallel PR review: security, performance, test coverage.