aldena learnupdated

what are the ai coding agents?

the main ones are claude code and openai codex in the terminal, github copilot coding agent and devin running asynchronously on issues, cursor and windsurf as editors, and open-source options like aider and opencode. the useful split is whether the agent works beside you or runs unattended on a task.

the landscape, split by how you use it

agentshapeworks best on
claude codeterminal, interactive or headlessmulti-file changes in a repo you know
openai codexterminal and clouddelegated tasks with a clear definition of done
github copilot coding agentassigned an issue, opens a pull requestsmall, well-specified issues
cursor, windsurfeditor with an agent modetight loop with a human in the chair
devinasynchronous, its own environmentlong-running scoped tasks
aider, opencodeopen source, terminalfull control, bring your own model

The row that matters is not the vendor. It is whether you are supervising continuously or checking back later.

why coding is the strongest agent domain

Because the environment tells the truth for free. The code compiles or it does not. The tests pass or they do not. The linter is not diplomatic. That gives the agent a feedback signal after every action, which is exactly the thing agents need and almost never have in other domains.

Compare it to a marketing agent, where nothing in the environment says whether the copy was any good. Same model, radically different reliability, and the difference is entirely the feedback loop.

where they still fail

  • Large unfamiliar codebases. Retrieval finds the wrong file and the agent confidently edits it.
  • Implicit conventions. Every repository has rules nobody wrote down, and the agent will violate them politely.
  • Tasks with an ambiguous specification. It will not ask. It will infer, and finish something.
  • Self-review. The agent that wrote the change is the agent judging the change, against the same assumptions that produced the bug.

The fourth is structural rather than a model weakness, and it is the one that gets worse as tasks get longer.

how to get real value from one

Small, specified tasks with a test to prove the result. Read-only until you trust it. Review every diff for the first month, because the failure mode is plausible and wrong rather than obviously broken. And write your conventions down somewhere the agent reads, since it cannot infer what only lives in your head.

how this works in aldena

Aldena runs coding agents as a team rather than one worker, which is the direct answer to the self-review problem. A project manager splits the work, engineer agents implement, and a separate reviewer reads the diff before it reaches you, arranged in an org chart you set.

Each agent works in an isolated room with its own server, connected to github or bitbucket through skills, and keeps memory of your default branch and conventions between runs, which is the fix for the implicit-conventions failure above. Nothing merges until you sign off, because approval is on by default.

You pick the model per agent from the published catalog, so the ranking between the tools above becomes a setting rather than a commitment.

ready when you are

spin up your first room.

one room per client, project, or product, staffed with a project manager, an analyst, engineers and a reviewer.