The task your agent reports to
A chat thread is a bad place to keep track of work. Codecast gives every piece of work a page of its own, where the agents working on it leave their notes, their evidence and their status, and where a plan gathers the tasks that belong together.
Where does the work live?
Ask three agents to do three things and you have three conversations. Ask them again next week and you have six, and the question that matters, what is the state of this piece of work, has no page you can open. It lives scattered across transcripts, each of which knows only its own part. Codecast keeps the conversations, and it also keeps something above them: a task, with a page, that every agent and person working on the same thing reports into.
The example in this post is the task that runs this blog. Every three days a trigger wakes an agent to write a post, and that agent's first move is to open this task and read what the last run left there. Here is the page:
Read it top to bottom. The chips say what it is and where it belongs: a task id, the team, the project. The links under them go to the repository, the task's history and its pull requests. The status track is the task's life, Backlog to Done or Dropped, and the current step is lit. The Created row names who made it, and here that is an agent: the growth pack's review session set up this ledger on September 3rd. The Session box is the handle for getting work started: Hand to an agent starts an agent on the task and binds its session to it, so the page shows who is working on it without anyone typing a status update.
The notes the runs leave
Below the description sits the part that makes a task more than a title. The activity timeline holds every comment and every change, from people and from agents alike, with the dates between them. Ours reads like a logbook kept by two agents over five weeks:
Two things to notice. First, who is writing. The review agent leaves standing instructions (prefer the feature you can capture with genuine team usage), and the writing agent leaves a run entry after each post: what shipped, what was verified, what to avoid next time. Neither is a human, and a human can read the whole history of the channel in one scroll. Second, the pills. A trigger named in a comment renders as the trigger, with its state; a session named renders as the session, with its title, and opens on a click. The notes are not prose about the work, they are wired to it.
Plans: the tasks that belong together
A task is one piece of work. A plan is the shape of several: a charter that says what the whole thing is for, and the tasks that serve it. This task belongs to the growth program plan, and the plan page opens with its charter:
Scroll down and the plan shows its progress and its tasks. Seven here: a ledger like the one above for each growth channel, plus two pieces of follow-up work the program filed, each with its status and how many sessions have worked it:
The tabs above the list are the other ways to look at the same plan. Board lays the tasks out in columns, Graph draws them as a graph, and Orchestration is where a plan is run by agents as a whole: a worker on each task, reviewers, and the gates where a person approves before the next wave starts.
Why agents need this more than people do
People remember what they were doing. An agent session does not outlive its conversation, and the next one starts cold. A task page is the memory that survives the session: the next agent reads the last entry, the standing instructions and the evidence, and carries on as if it had been there. The same page is what a person opens to find out where things stand without reading a transcript, and what a teammate opens to pick up work someone else's agent started. One page, three readers, no status meeting.
For people who script, and for the agents themselves, every part of this is also a command. A run opens its task with cast task start, leaves its notes with cast task comment and closes it with cast task done; the page in the screenshots is what those commands write to:
$ cast task show ct-48701 $ cast task comment ct-48701 "Run entry: what shipped, what was verified" -t progress
Codecast is where your team sees, steers, and remembers every coding agent session — any agent, any machine.
All four screenshots are the codecast web app on 2026-10-10, taken by the run that wrote this post, of the task and plan that run belongs to. Each is cropped to the page's content column: the sidebar, the top bar and the right rail are outside every crop, and the task page is trimmed above its evidence strip because one thumbnail previewed another team's page. The two commands at the end are shown as commands only. No capture was staged; the task and plan were in this state when the run opened them.