Agents in the inner loop, agents in the outer loop
A recent All-Hands blog contrasts AI agents operating in the inner loop—local IDE or CLI tools that accelerate incremental coding—with emerging outer loop agents that run autonomously in the cloud as part of CI/CD, Slack, Jira, or GitHub workflows. Published December 2, 2025, the piece explains why teams will likely adopt both modes: interactive agents for ad hoc developer work and cloud agents for large-scale, repeatable maintenance.
Inner and outer loops defined
The inner loop describes the familiar edit-run-debug cycle on a developer’s laptop: small changes, fast feedback, personal tool choice (neovim, VS Code, Zed, Cursor, etc.). The outer loop begins after a git push—CI/CD, code review, issue tracking—and requires shared tooling across a team.
Inner-loop agents: pair-programmer assistance
Agents in the inner loop run locally inside an IDE or as a CLI. They offload rote implementation tasks while the developer focuses on design and problem solving. Research referenced in the original discussion suggests roughly a 40–50% productivity gain in implementation when developers work with LLM-powered assistants (see research links below), though studies also report increases in debugging time. Local agents remain interactive tools that typically require human supervision for risky operations.
- IDE integration: https://docs.openhands.dev/openhands/usage/run-openhands/acp#acp-integration
- CLI mode: https://docs.openhands.dev/openhands/usage/run-openhands/cli-mode
- Productivity studies: https://arxiv.org/abs/2507.08149 and https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5713646
- Increased debugging time study: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
Outer-loop agents: autonomous, isolated, and scalable
Outer-loop agents run on their own infrastructure—GitHub Actions, Kubernetes pods, or managed services such as https://app.all-hands.dev/—with a dedicated filesystem, terminal, browser, and copy of the codebase. Two attributes stand out: safety and scalability.
Safety
Running agents in isolated environments limits blast radius. Outer-loop agents access only the tools and credentials explicitly provided, which enables unattended execution with far less risk than fully privileged local runs. This containment makes it practical to let agents operate autonomously for routine tasks, provided that sensitive keys and permissions are managed carefully.
Scalability
The cloud removes the practical ceiling of 2–3 concurrent local agents. Containers or VMs let teams spin up dozens or thousands of agents in parallel—useful for organization-wide work that touches many repositories. The trade-off is cost, but the technical ability to scale is straightforward.
Practical targets: tech debt and toil
Outer-loop agents are especially well suited to repeatable, high-volume maintenance that otherwise consumes engineering time.
-
CVE remediation: Vulnerability discovery and patching is a procedure-heavy workflow: locate affected packages or Dockerfiles, determine safe updates, apply changes, and verify tests. Outer-loop agents can research, implement, validate, and open PRs autonomously across many repositories—similar in spirit to Dependabot but with broader automation scope. Further details: https://openhands.dev/blog/remediating-cves-at-scale-with-agents
-
Error-log remediation: Scanning Datadog logs for new error patterns and mapping incidents back to the responsible repository is another area for automation. Agents can propose fixes—improving error handling or downgrading noisy errors—so teams wake up to PRs or proposed patches rather than being swamped by toil. See: https://openhands.dev/blog/debugging-production-issues-with-ai-agents-automating-datadog-error-analysis
Complementary roles: both loops matter
The most effective workflows combine both approaches. Inner-loop agents are ideal for exploratory feature work and tight feedback cycles inside an IDE. Outer-loop agents excel at high-volume, repeatable maintenance tasks that benefit from autonomous, isolated execution at scale. Teams may adopt different inner-loop tooling per developer while standardizing on an outer-loop platform or SDK (for example, https://github.com/OpenHands/software-agent-sdk).
Adopting outer-loop automation can reduce backlog and free engineers for higher-value work—provided that credential management and governance are handled deliberately.
Original article: https://www.all-hands.dev/blog/20251202-agents-in-the-outer-loop
