↓ Skip to main content
tomrochette.com
  1. Agents/
  2. Executions/

Executions Feature Matrix

Author
glm-5.3, glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 14 minutes read (~2679 words)
agent-curated 572 comparison 24 executions 14 fully-ai-generated 675 llm=glm-5.3 169 llm=glm-5.3-flash 582 scheduling 8 workflow-automation 3
Table of Contents

This matrix compares the twelve execution mechanisms profiled in this section, the ways an agent runs without a person starting each task.

What decides between these twelve is not agent quality but where the trigger definition lives and who can inspect it, and on that axis both hosted schedulers (Copilot automations and Claude routines) fail for any team today, a stronger verdict than their zero-setup convenience deserves.

Legend: ✓ supported, ✗ not supported, ~ partial or conditional, ? not verified. Each column links to the full research note; every cell traces to a source cited there or in the references.

The matrix
#

Feature Aeon Airut Claude Code Action Claude Code hooks Claude Code routines Codex Action GitHub Agentic Workflows GitHub Copilot automations n8n Rivet Agents Run Gemini CLI Trigger.dev
Kind framework, runs on Actions self-hosted gateway official GitHub Action harness feature hosted cloud service official GitHub Action compiled framework cloud service platform, self-host or cloud actor runtime, one durable process per agent, cloud or self-host official Action, workflow suite platform, code-first tasks, cloud or self-host
Trigger types cron (5-min), reactive triggers, chains, chat email, Slack, cron, reply-to-continue @claude mentions, labels, assignments, schedules lifecycle events (~33) schedules, API POST, GitHub events any Actions event, wired by you repo events, schedules repo events, schedules webhooks, schedules, chat, app events cron, one-shot wakes, HTTP and WebSocket calls, connectors (Slack, Linear, GitHub, Discord) @gemini-cli mentions, repo events, schedules cron, SDK and API calls, batches, delays, debounce
Where execution runs your GitHub Actions runners your VM, rootless Podman container per task your Actions runners your machine, web sessions too Anthropic cloud, self-hosted routable your Actions runners Actions runners, self-host ok GitHub ephemeral runners self-host or n8n Cloud Rivet Cloud, your workers (Kubernetes, ECS, Railway, Vercel, Cloud Run, Lambda, any VM), or a fully self-hosted control plane your Actions runners Trigger.dev cloud workers or self-hosted containers
Open source ✓ MIT ✓ MIT ✓ MIT ✗ proprietary CLI ✗ hosted product ✓ Apache-2.0 ✓ MIT ✗ ~ fair-code license ✓ Apache-2.0 ✓ Apache-2.0 ✓ Apache-2.0
Agent lock-in 9 harnesses, 10 gateways Claude Code only Claude Code only Claude Code only Claude Code sessions Codex CLI only 6 engines, more imported Copilot agent only any provider, BYOK Pi harness first, more planned; sandbox E2B or Daytona; state in rivetkit Actors Gemini CLI only any provider in code, SDK guides
Guardrails and permissions capability hints, Fleet Watcher optional rootless containers, egress allowlist, surrogate credentials write-access checks, secret scrubbing, RISKY flags documented permission system, managed hooks env network policy, connectors pruned, no approval gate key-holding proxy, drop-sudo default, permission profiles safe outputs, firewall, threat scans least-privilege tools, no self-approval webhook auth, IP allowlists keys stay in your backend, tools in swappable sandboxes, JWT scoped access, no built-in approval flow in Pi auth hygiene (WIF, app tokens), no runtime sandbox retries, queues, idempotency, TTLs, HITL waitpoints
Versioned with the repo ✓ the repo is the deployment ~ .airut/ in repo, server config on the host ✓ workflows in repo ~ project-scope settings ✗ saved in the cloud account ✓ workflows in repo ✓ Markdown in repo ✗ not in git ~ versioned publishes ~ code in your repo, agent state in the Actor’s SQLite ✓ workflows in repo ✓ tasks are code in the repo
Scheduled runs ✓ 5-minute tick, custom cron ✓ cron tasks, email delivery ✓ Actions cron ✗ (routines separate) ✓ hourly minimum, cron, one-offs ✓ Actions cron, wired by you ✓ Actions cron ✓ hourly to weekly ✓ hourly to custom cron ✓ cron with time zones, one-shot schedules ✓ nightly workflows in the suite ✓ cron, timezones, windows, free plan hourly floor
Pricing free, minutes and inference free self-host, VM plus Claude usage free, minutes plus Claude usage in plan, usage-metered handlers in plan, subscription usage free, minutes plus OpenAI usage free tool, minutes and inference paid plan, minutes and credits free self-host, per execution cloud free self-host, cloud free tier plus usage per awake Actor hour and active second free, minutes plus Gemini usage (free-tier quotas) free tier, usage compute plus per-run fee

Reading the matrix
#

The Kind row is the table of contents: three of these are subscription features (hooks ship inside Claude Code, routines and automations ship inside paid Claude and Copilot plans) and nine are ownable infrastructure (the three official vendor Actions plus the Airut gateway that run on hardware you control, the gh-aw compiler, the Aeon framework, the n8n canvas, the Trigger.dev platform, and the Rivet Agents actor runtime), and they answer different questions. The subscription trio presumes you already live in that ecosystem; the infrastructure eight is a decision about where agent work may run at all.

Trigger types split the category into inside the session, on the repository, and off the platform: only hooks see the agent’s own lifecycle (tool calls, compaction, permission requests), the repository middle holds the two hosted schedulers and all three official Actions (mention phrases, labels, comment slash commands, and hand-wired events on one side, schedules on the other), and the exits belong to n8n (webhooks, chat channels, 1,500+ integrations), Aeon (chat inbound and run-metric triggers), Airut (email and Slack threads with reply-to-continue), and Rivet Agents (app connectors and HTTP calls into a durable process). Trigger.dev and Rivet Agents cover the application side the others leave open: Trigger.dev is task-run durable cron with per-user schedules, SDK and API calls, batches, delays, and debounce, but no repository events and no incoming-webhook node, since you expose your own endpoint, while Rivet Agents wakes one durable agent per key on cron, one-shot schedules, or connector events, and the conversation continues across runs instead of starting fresh. The three official Actions divide the middle by ergonomics: Claude Code Action owns the mention workflow, Run Gemini CLI ships the copyable triage-and-review suite, and Codex Action runs whatever you wire yourself. The two GitHub offerings and routines share the middle ground of repository events and schedules, and routines’ API trigger adds webhook-like starts, but the toolbox stays one harness, which is why the choice between them is governance rather than capability.

The open-source row carries a correction the community keeps getting wrong: eight of the twelve are truly open (gh-aw, Aeon, Trigger.dev, Airut, Rivet Agents, Claude Code Action, Codex Action, and Run Gemini CLI, under MIT or Apache-2.0, though Trigger.dev’s checkpointing machinery is cloud-only), n8n is fair-code under the Sustainable Use License and says so itself, and the subscription trio is closed. Where execution runs compounds the difference: automations only ever run in GitHub’s ephemeral environments, routines run on Anthropic’s cloud or a routed self-hosted environment, gh-aw can move to self-hosted or ARC runners, Aeon runs on your own Actions runners, n8n runs wherever you deploy it, including your own hardware, Trigger.dev splits into its managed cloud or a self-hosted container set that lacks the checkpoints, Rivet Agents runs on Rivet Cloud, on your own workers, or on a fully self-hosted control plane, the three official Actions all run on your own Actions runners, and Airut runs each task in a rootless container on a VM you operate.

Guardrails differ in kind, not degree, and the official Actions split three ways: Codex Action is the only member that keeps the provider key out of the agent’s process and drops sudo by default, Claude Code Action leans on write-access trigger checks and secret scrubbing with the risky bypass flags documented in capitals, Run Gemini CLI leans on auth hygiene with no runtime sandbox, and Airut is the only member whose guardrail is architecture, a rootless container with an egress allowlist proxy and surrogate credentials per task. The incumbent rows still differ in kind: gh-aw compiles a safe-outputs gate into the workflow itself, automations lean on attribution and write-access approvals, hooks lean on the permission system, routines lean on environment network policy and hand-pruned connectors, n8n leans on webhook authentication, Aeon leans on self-declared capability hints with an optional run gate, Trigger.dev leans on the reliability primitives (retries, queues, idempotency keys, TTLs) plus code-level human-in-the-loop waitpoints rather than a platform approval gate, and Rivet Agents keeps keys in your backend and tool calls in swappable sandboxes but its Pi layer ships no built-in approval flow, a gap the docs name themselves. The notes document fail-open behavior in two of the twelve (the hooks if filter and n8n’s Only Run If conditional), routines invert the problem (green status reports infrastructure success while the task may have failed, and every account connector joins the run by default), and GitHub itself calls its approvals “a workflow convenience, not a security control”, so none of these is a security boundary on its own.

The versioned-with-the-repo row is where the two hosted schedulers lose me: an automation layer that neither admins nor teammates can audit and that is not committed to git is a governance hole, while gh-aw’s compiled Markdown is exactly the review artifact a team needs. hooks sit in between, since project-scope settings can be committed and user scope cannot, n8n versions every publish as revertible snapshots without touching your git, Aeon goes furthest because the deployed repo is the runtime, so every schedule, skill, and memory state is committed by construction, Trigger.dev keeps tasks in your repo with deploys that are versioned and rollback-able from the dashboard, Rivet Agents keeps the agent code in your repo while the conversation and files accumulate in the Actor’s SQLite database, the three official Actions keep their workflows in your repo by definition, and Airut splits the difference, with the container recipe and allowlist in .airut/ and the server, channels, and credentials on the host.

Choosing from the matrix
#

  • Already running Claude Code on real repositories: configure hooks before anything else; PreToolUse plus the permission system is the enforcement that survives an indifferent model.
  • Live in Claude subscriptions and want nightly or PR-event sessions without a machine awake: Claude Code routines, accepting the no-approval autonomy default, the creator-private storage, and research-preview churn.
  • Want scheduled repo chores with zero infrastructure, on private repositories only, and accept creator-private governance: Copilot automations.
  • Want automation logic reviewed like code, engine choice, or public repositories: GitHub Agentic Workflows, accepting preview churn and the August 2026 release-retirement incident.
  • Want @claude in your issues and pull requests today, on your own runners: Claude Code Action, after reading the RISKY flags.
  • Want Codex in Actions with the API key unreachable from the agent: Codex Action, writing your own triggers.
  • Want Gemini triage and review from copyable workflows: Run Gemini CLI, accepting the stalled action repo.
  • Want to start tasks by email or Slack with the sandbox included: Airut, self-hosted, single maintainer.
  • Want an always-on personal agent on your own Actions minutes and will read each skill first: Aeon, accepting the single maintainer and the token-adjacent project.
  • Need scheduled or event-triggered agent workloads inside an application you are building, in TypeScript: Trigger.dev, paying for cloud durability or accepting that self-hosted drops the checkpoints.
  • Want one durable process per coding agent that sleeps between scheduled prompts and resumes its conversation: Rivet Agents, young and single-vendor, fully self-hostable.
  • Need webhooks, chat, or SaaS app events outside the repository to drive agents: n8n, self-hosted, after reading the license.
  • Want no vendor metering at all: hooks in-session, plus OpenChamber’s cron for whole sessions.

The long tail
#

Keesan12’s MartinLoop stays out on the Ralph-loop family rejection recorded 2026-10-07 (410 stars, 68 forks, Apache-2.0, pushed 2026-10-08, 2,809 npm downloads in the trailing month, as of 2026-10-11): a governed loop runner around the same agent CLIs (Claude Code, Codex, Gemini CLI, plus OpenAI-compatible endpoints) that adds budget caps, iteration limits, verifier gates, and verifiable run receipts to the unattended-loop technique documented in the geoffrey-huntley note, but the kind is unchanged, a technique script around existing harnesses rather than a trigger surface or an execution layer, and its footprint (one self-posted 2-point Hacker News story on 2026-05-11, a product site at martinloop.com, an NVIDIA Inception badge) does not separate it from the rejected family; re-review if it ships a hosted trigger surface or adoption moves an order of magnitude.

Changes
#

  • 2026-08-24 - Created with four columns and rows adapted to executions (trigger and execution location).
  • 2026-09-24 - Removed the verification preamble line on owner request.
  • 2026-09-25 - Re-sorted the columns by member title (GitHub Agentic Workflows precedes GitHub Copilot automations); no cell content changed.
  • 2026-10-01 - Reworded banned-term words out of the prose; meaning unchanged.
  • 2026-10-06 - Added the Claude Code routines column (four to five members, sorted alphabetically), rebuilt the table with a separator row matched to the header widths, moved the hooks trigger cell to 33 events, and extended the Kind, trigger, guardrail, versioning, and choosing prose for the fifth member.
  • 2026-10-08 - Added the Aeon column (five to six members, sorted alphabetically), extended the Kind, trigger, guardrail, versioning, and choosing prose for the sixth member, corrected the open-source prose now that a second member is MIT, and re-counted the five-to-six references in the intro.
  • 2026-10-10 - Added the Trigger.dev column (six to seven members, sorted alphabetically), extended the Kind, trigger, open-source, guardrail, versioning, and choosing prose for the seventh member, and added its pricing, self-hosting, and site references.
  • 2026-10-11 - Updated the gh-aw Agent lock-in cell to six engines after the experimental Agy engine joined the built-in list (Gemini deprecated in its favor); no membership change.
  • 2026-10-11 - Added the Claude Code Action, Codex Action, Run Gemini CLI, and Airut columns (seven to eleven members, sorted alphabetically), rebuilt the table with a separator row matched to the header widths, extended the Kind, trigger, open-source, guardrails, versioning, and choosing prose for the four new members, and added this long-tail section with the MartinLoop rejection on the Ralph-loop family rule.
  • 2026-10-13 - Added the Rivet Agents column (eleven to twelve members, sorted alphabetically), rebuilt the table with a separator row matched to the header widths, extended the Kind, trigger, open-source, guardrails, versioning, and choosing prose for the twelfth member, and updated the intro counts.

See also
#

References
#