ABCA
Autonomous background coding agents on AWS: submit a task, get a ready-to-review pull request — no babysitting, no IDE sessions.
- Stars
- ★ 161
- Last updated
- 2d ago
- License
- MIT-0
- Primary language
- TypeScript
- FA score
- 74/100 · Some gaps
At a glance
- How it runs
- Works with
- Platform-specificClaude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- High · needs real infrastructure
- You'll need
- Typical use
- Teams that want routine coding work (bug fixes, small features) handled in the background so engineers can focus on design and decisions
- Not a fit if
- Teams without an AWS account or unwilling to run workloads on AWS
- Developers who need real-time conversational interaction instead of async task submission
- Teams seeking a production-ready solution — the project is explicitly experimental and educational
- Source review
- 74/100 · Some gaps
What does this agent do, and when should you use it?
ABCA (Autonomous Background Coding Agents on AWS) is an official AWS sample repository, aws-samples/sample-autonomous-cloud-coding-agents, demonstrating what a self-hosted background coding agents platform looks like on AWS. You submit a coding task via Slack, Linear, Jira, CLI, or webhook, and an agent autonomously clones the target GitHub repository, creates a branch, writes code, runs tests and lint, and opens a pull request from an isolated MicroVM sandbox. Execution follows a hybrid blueprint workflow that mixes deterministic steps (admission, pre-flight checks) with LLM-driven agentic steps, with each task running in an ephemeral, asynchronous, repository-scoped environment. A memory system captures PR feedback to improve future runs, and the platform ships with per-task budgets, concurrency limits, and blast-radius enforcement. The current version is in maintenance mode with a V2 in development, and the project is explicitly experimental and educational — not intended for direct production use.
Accepts task submissions from Slack, Linear, Jira, CLI, or webhooks. An orchestrator executes a blueprint with five phases: 1) Admission — validates the request, checks concurrency limits, and queues the task; 2) Context hydration — gathers the task description, GitHub issue body, repo-intrinsic knowledge (CLAUDE.md, README), and memory from past tasks on the same repo; 3) Pre-flight — fail-closed readiness checks (e.g., GITHUB_UNREACHABLE, REPO_NOT_FOUND_OR_NO_ACCESS) that fail doomed tasks fast before consuming compute; 4) Agent execution — the agent runs in an isolated MicroVM with persistent session storage, cloning the repo, creating a branch, editing code, committing, and running tests and lint; 5) Finalization — infers the result (PR created or not), runs optional validation (lint, tests), extracts learnings into memory, and updates task status. Deployment is via AWS CDK, and the repository also ships a Claude Code plugin (docs/abca-plugin) providing six skills (setup, deploy, onboard-repo, submit-task, troubleshoot, status) plus two specialized subagents, cdk-expert and agent-debugger.
- Teams that want routine coding work (bug fixes, small features) handled in the background so engineers can focus on design and decisions
- Development teams using Slack, Linear, or Jira as task sources who want those tools to trigger automated coding flows
- Engineering organizations that want tasks executing around the clock to cut cycle time instead of queueing behind a human's calendar
- Platform/infrastructure teams that want to onboard their own GitHub repositories into an automated PR pipeline that learns per-repo over time
- Architects studying how to build governed, observable autonomous agent platforms on AWS
How do you install or deploy this agent?
The recommended path is the repository's bundled Claude Code plugin, which provides guided setup workflows. Clone the repo and launch Claude Code:
git clone https://github.com/aws-samples/sample-autonomous-cloud-coding-agents.git
cd sample-autonomous-cloud-coding-agents
claude --plugin-dir docs/abca-pluginThe plugin's skills and commands become available immediately. If you use Claude Code via VS Code or JetBrains, add --plugin-dir docs/abca-plugin to the extension's CLI arguments setting.
For manual installation, install [mise](https://mise.jdx.dev/getting-started.html) first (set MISE_EXPERIMENTAL=1 for monorepo-prefixed tasks), then follow the Developer Guide (docs/guides/DEVELOPER_GUIDE.md) to prepare your environment and deploy the application to your AWS account via CDK. Specific AWS permission configuration and LLM key details are covered in that guide.
How do you use this agent?
With the Claude Code plugin, trigger skills in natural language: say "get started" for guided setup, "deploy" to run the CDK deployment, "add a repo" to onboard a new GitHub repository via the Blueprint construct, "submit task" to submit a coding task (with prompt-quality coaching and a quick mode), or "status" for a platform health check.
Manually, use mise tasks to install and build:
mise run install
mise run buildMonorepo-prefixed tasks (e.g., mise //cdk:build) require MISE_EXPERIMENTAL=1.
Once deployed, submit tasks through Slack, Linear, Jira, CLI, or webhook, then wait for the agent to finish in its isolated environment and open a PR for you to review.
What are this agent's strengths and limitations?
- End-to-end autonomous loop: from task submission to a reviewable PR with no human intervention — the agent clones, codes, tests, and opens the PR itself
- Hybrid blueprint workflow separates deterministic steps (no LLM, cheap, predictable) from LLM-driven steps for cost control
- Built-in governance and observability: per-task budgets, concurrency limits, blast-radius enforcement, and fail-closed pre-flight checks that fail doomed tasks fast
- A memory system captures PR feedback and improves future runs, while each task runs in an ephemeral MicroVM with no in-process state carryover
- Ships a complete Claude Code plugin (six skills, two specialized subagents, a SessionStart hook) with guided deployment and troubleshooting
- Explicitly experimental and educational — not intended for direct use in production environments
- Deeply tied to AWS: requires an AWS account and a full CDK deployment, making setup genuinely infrastructure-heavy
- The current version is in maintenance mode with V2 in development, so a future migration is signaled (none required yet)
- The asynchronous fire-and-forget model does not fit workflows needing real-time conversational interaction
- May collect anonymous operational metrics (deployment region, construct name and version, related information)
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| ABCA This agent | 74 · Some gaps | Self-hosted serviceFree + model costs | ★ 161 | 2d ago | TypeScript | Claude Code |
| Remote SWE Agents | 66 · Some gaps | Self-hosted serviceFree + model costs | ★ 244 | 13d ago | TypeScript | — |
| Open SWE | 68 · Some gaps | Self-hosted serviceFree + model costs | ★ 11k | today | Python | OpenAI API |
| Smart Ralph | 73 · Some gaps | Agent plugin / skillFree + model costs | ★ 559 | 1mo ago | Shell | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Least privilege is exceptionally well evidenced: CI workflows use explicit deny-all top-level permissions, job-level minimal grants, SHA-pinned third-party actions, and the auto-approve workflow explicitly excludes fork PRs and pairs with Mergify dismiss-stale rules to block unreviewed merges. Attachment downloads enforce SHA256 checksums, read-only file modes, and cleanup on failure; agent runs in isolated MicroVMs. Deductions: operational telemetry (region, construct name/version) is enabled by default with no visible opt-out; user confirmation is limited to post-hoc PR review with no in-flight human gate.
Failure handling is excellent: fail-closed pre-flight checks (GITHUB_UNREACHABLE, REPO_NOT_FOUND_OR_NO_ACCESS) fail fast with explicit reason codes; the test conftest shows a deeply engineered hang watchdog and session-cache leak guard; self-mutation detection fails the build. Deductions: dependency availability is weakened by a resolutions block pinning many transitive dependencies and a mandatory experimental mise flag (MISE_EXPERIMENTAL=1).
Trigger precision is a standout: the plugin skills table lists explicit trigger phrases, a SessionStart hook advertises capabilities, and auto-approve fires only on labeled events from non-fork PRs. Audience is clear (self-hosted background coding agent platform), and boundaries are conveyed via the blueprint's deterministic/agentic split. Deductions: hard operational boundaries on what the agent may not modify are not shown in the provided evidence.
Information architecture is excellent: README → ARCHITECTURE.md → guides → docs site; MIT-0 license matches the LICENSE file; maintenance responsibility is explicit (maintenance mode, V2 direction, issue priority-label process); limitations are honestly declared (Experimental badge, non-production disclaimer). Deductions: package.json shows version 0.0.0 and private with no changelog; the LICENSE copyright year reads 2026, which is anomalous; prerequisites (mise experimental flag, AWS account, CDK) are scattered in guides with only pointers in the README.
Output usability is strong: the deliverable is a reviewable PR with lint/tests/policy checks built in; cost controls (per-task budgets, concurrency limits, blast-radius enforcement) are claimed. Deductions: quantified benefit is asserted only — the 'learns over time' memory system has no benchmark or evaluation evidence in the provided files, so marginal value over direct Claude Code use cannot be substantiated.
Most claims are traceable: architecture claims point to ARCHITECTURE.md, CI behavior is directly checkable in workflow files (permission matrices match the explanatory comments), and test files substantiate the attachment integrity mechanism. Facts and inferences are largely separated (explicit experimental disclaimer). Deductions: cross-source corroboration is limited — security and learning-effect claims lack independent evidence in the supplied files, and the publisher is registry-unverified, capping this at medium under static review.
- Anonymous operational metrics (region, construct name/version) are collected by default; compliance-sensitive deployments should verify opt-out availability and data scope before deploying.
- The project is in maintenance mode with team investment moved to V2; do not make long-term production commitments — the disclaimer explicitly states it is not for production use.
- The auto-approve + Mergify auto-merge path has anti-abuse constraints, but reproduce the label and branch-protection configuration fully in your own org before adopting, to avoid any unreviewed-merge path.
- The agent autonomously clones repositories, commits code, and opens PRs; audit GitHub token scopes and the repository onboarding allowlist yourself.
- Installation depends on a mise experimental feature (MISE_EXPERIMENTAL=1), a non-stable toolchain prerequisite.
- package.json is at 0.0.0 with no changelog; upgrades and security fixes are traceable only via git history and issues.