AI agent adoption · making it stick
AX for everyone
Orchestrate your AI agents as one team, with Marblo We help you decide what to automate, set up controls with Marblo, and build the capability to run it yourselves.
47 posts · 42 open-source projects listed

Notes from our own runs
Where to restart when a build gets stuck.
- Marblo Build Notes 10 — Why Telegram Sending Failed While Receiving WorkedTelegram messages arrived, but sending failed with ETIMEDOUT. Trace the IPv6 path, IPv4 connection latency, and Node's connection attempt timeout, with a diagnostic sequence builders can reuse.
- Marblo Build Notes 11 — Merged, So Why Isn't It DONE?A merge records that code was integrated. How Marblo verifies the combined result, keeps a signed record, and checks remaining work before completion, plus the human cost of a conservative gate.
- Is a More Complex Agent Harness Better? Separate Cost from Parallel WorkCompare a fixed-model harness study with Agensh's 1,024-agent experiment. Understand their different conditions and measure quality, cost, time, and integration work before adoption.
What your team faces now
You already use AI coding tools. But everyone picked their own, and the company sees none of it.
- Nobody knows who uses which tool, or how much. No one approved it, yet everyone uses it.
- Agent changes merge unreviewed, or nobody dares use them in production. There is no middle ground.
- The output survives, the reasoning does not. When someone leaves, the next person starts over.
- You learn each team's AI spend when the invoice arrives. There is nothing to budget from.
- Nobody inside can assemble what a security review needs, so the review never starts.
Isn't Claude Code enough on its own?
Usually, yes. Claude Code and Codex are good tools, and plenty for one task at a time. We use them that way too.
Trouble starts when one becomes two, then five. Past that, tracking the agents takes longer than running them. Here is where it gets stuck.
- Only one in view at a time
More windows, more agents. By the third, you lose track of what finished where. Each new window is one more thing to miss.
MARBLO
Everything on one screen. You read state, not a chat log.
- They overwrite each other
Two agents in one repository overwrite each other. You end up running them in turn, which defeats parallelism.
MARBLO
One git worktree per ticket. They physically cannot touch each other's files.
- Nothing stops a merge
Review relies on memory and gets skipped when busy. A policy is a document, not a flow.
MARBLO
Work must pass REVIEW to merge. The flow enforces the policy, not a document.
- No model choice per task
Claude Code is Claude. To use Codex here and Grok there, you must switch tools. You end up tied to one vendor.
MARBLO
Set a model per ticket. If one vendor has trouble, work continues.
- Cost shows up on the bill
No live view of who spent how many tokens on what. You cannot budget or catch an overrun.
MARBLO
Usage and spend land in the workspace by vendor, model, and agent, down to which model each agent actually ran.
- The reasoning disappears
The code survives; the why leaves with the chat window. Months later you re-derive it from scratch.
MARBLO
An append-only record per ticket: which agent changed what, and on what grounds.
Agents are cheap to start. Knowing what the fleet is doing is the hard part.
Marblo official repository
No control, no adoption
Running agents in parallel is easy. The hard part is knowing what changed while they run, and keeping a human on final approval. That, not performance, is what stalls adoption reviews.
feature/audit-panel
ahead 3 · behind 0 · no conflicts
fix/board-query
ahead 7 · behind 4 · 2 conflicts
archive/repo-connect
ahead 0 · behind 12 · merged
- Isolated execution
- One git worktree per agent. Parallel runs never overwrite each other, and a conflict shows as a conflict.
- Approval gate
- Finished work goes to REVIEW and merges only on a person's approval. Review is not left to good intentions: without approval, nothing moves on.
- Audit log
- Team account activity is available on the organization page. Audit logs are agreed with you under Enterprise.
- Cost tracking
- Subscription and AI usage are separate, with consumption per model and agent. You budget from real numbers.
Not only about code
Split a goal, hand each branch to a model suited to it, and have a person merge the result. That way of working is not specific to code.
Marketing, data analysis, GA4 setup, and planning docs run on the same structure, so teams without engineers can use it too.
마블로 클로즈드 루프 공개! AI 에이전트가 다이노 게임을 자동 완성했습니다
Actual run recording · Korean video · channel · 데이터가 답이다 · YouTube


Models in rotation
Choose models for each task, starting with Claude, Codex, and Grok. Connect other models through BYOM.
You can leave with what you built
A real worry in adopting a tool is whether you can leave later. Keep your process in one tool and you are stuck if it gets pricey or changes course.
173 skills, MCP servers, agents, workflows, and knowledge packs are MIT-licensed in the official repository. They are plain files that run in your current CLI, so they stay if you drop the app. 22 are for Korean work: DART filings, real-estate prices, case-law search, HWP editing.
It does not fit every team
Marblo suits teams that delegate several tasks at once and review the results. Compare the situations below with your team.
Good fit
- Teams where one person keeps several tasks moving at once
- Organizations requiring human review before merge
- Organizations that must explain AI spend and work history per department
- Organizations that see single-vendor lock-in as a risk
- Teams that could also apply it to non-engineering work (marketing, analysis, planning)
Not yet
- Work small enough that one person working in order is faster
- Organizations that cannot yet budget for AI usage separately from subscriptions
- Linux-only environments. Only macOS and Windows are supported today
- Teams that need process before tooling. We address that with training first
- When the goal is cutting headcount right away. It is not built for that.
Adoption is not one step
Turning on a subscription is not the same as the team using it. Each stage has its own decision; stopping at any can be right.
- 01
Free first assessment
The free first assessment helps determine whether now is the right stage to adopt: 1–2 hours and a one-page summary.
- 02
AX assessment
Pick workflows that suit agents, the first pilot scope, and the success metrics.
- 03
Team training
Practice task splitting, approval criteria, and failure handling so your team can direct agents.
- 04
Pilot
One team runs Marblo on real tickets for four weeks. Measure the results and decide on licenses.
- 05
Marblo licenses
You roll out the proven approach company-wide. We supply Marblo seats and the product.
AX assessment, team training, and pilots are quoted after consultation. Credit terms are explained with the quote. Marblo licenses follow public pricing on marblo.app.
Where would you like to start?
Start with company adoption, developer resources, trying Marblo, or becoming a partner.
- Companies
- Evaluating adoption? Start with a free first assessment: 1–2 hours and a one-page summary.
- Builders and developers
- Want to build and read first? Start in the resource library: public resources and notes from our own runs.
- Trying Marblo yourself
- Want to try it first? Start with Marblo Founders.
- Resellers and partners
- Already pitching AI agents to clients? We can work together.