dFlow is changing.
We’re moving from open source to an always-free tier on dFlow Cloud.
That change gives us room to focus on what matters after the first deploy: making production infrastructure simpler to use, operate, and scale.
We’re also building a Kubernetes-based clustering layer. The goal is simple: deploy services without managing the underlying compute or handling scaling yourself.
Same product surface. Less infrastructure work between you and production.
Read what changed and why: https://t.co/hZLmZ4xWMk
#dFlow #DevOps #Kubernetes #CloudDeployment
dFlow build log #1: the first deploy
Before features, we needed one path that works.
You log in from the CLI.
You make a project and an environment.
You give dFlow a container image.
dFlow deploys it to Kubernetes and keeps it running.
Every later feature builds on this path.
@samir_sawarkar Our guess: the agent copies the pattern it saw most often. Most short tutorials put the query in the route handler.
Write the layer rule in a file in the repo, and the agent will follow it.
@dwsobest The same test works for deploys. Before an agent deploys a change, it must know how to undo it.
Speed is safe only when the rollback is ready.
@Prayag12B We think the context must live in the repo, next to the code. Each team owns one file for its part. The agent reads all of them before it acts.
If the context is not in git, it gets old fast.
@bendee983 Agreed. Agents make new code cheap, so each line you keep has a higher cost to read and maintain.
Often the best change is the one that removes code.
@notesbyfloat A linter checks your code against rules that you chose. Vibe coding accepts code that you did not read.
The difference is review, not the tool.
@satyam_builds@odd_joel Both, because they catch different failures. A spend cap limits the cost after the action. An approval step stops the action before it starts.
For any write that costs money, we'd use both.
@Mkandeger2@odd_joel Agreed. We think "off" must be part of the same action. When the agent starts a paid feature, the system sets an end time at that moment.
Then the agent has nothing to remember.
@odd_joel This is the real risk when agents control infra. The agent starts a paid feature, and nothing stops it later.
We think each agent change needs a log entry and an end time.
Did you get an alert, or did you find it on the bill?
@_kanaidu Good rule. The team gets one hour to answer first. If nobody answers, the agent does the rollback.
Next question: what if the deploy had a database migration? Would you let the agent undo that too?
A deploy fails at 3 AM. Your agent sees the errors, and it knows the last good version.
Should it do the rollback on its own? Or should it wait for a human?
Our view: yes, but only to a version that passed before.
Which actions would you let your agent do alone?
@theomlondhe Good find. A plugin can change an agent's defaults, and the user may not see it. We think every default must be in a config file that the team can read.
@saiitoshii A benchmark from real shipped tasks is a smart idea. Public benchmarks don't know your codebase. We'd add one more test: can the agent deploy the change and fix a failed build?
@kedaya We see the same trend. Agents need a cloud that they can use with no human help. For us, the hard part is the deploy step. The agent must see why a build failed, then fix it or do a rollback.
@samir_sawarkar Our guess: Alpine uses musl, not glibc. Many Python and Node packages then need extra build steps, or they fail. Debian slim is the safer default for an agent that must ship without help.
@techyoutbe A green pipeline and Ready pods aren't the same as a good release. Rollback only helps if something is watching the app, not just the rollout. The hard question is the one you asked: how do you notice the failure when the deploy itself succeeded?
@ti_codes The code was clean and staging still went down. The gap isn't better syntax. It's the agent seeing the pool run out as it happens, not after someone spends Friday night in CloudWatch.
@vineerpasam If the agent's whole job is "did the deployment crash?", the answer is only as good as what it can see. A passing pipeline isn't a report from the running app. Curious what it had access to: logs, events, or just the CI status?