agent-os adapted task builder [bert v0.0.1] (was an experiment)

[ssr1ram]

edited Oct 20

After a number of iterations, I now have adapted agent-os workflows and taken inspiration from the agent-os approach and integrated a task manager (inspired by backlog.md, taskmaster.ai) . So, now I can

• plan product, create a new spec, iterate on it, generate tasks, execute tasks - all in a granular manner with no implicit dependency yet which inherits context as needed.

other points of note

• for small one of tasks i can create a task and execute it
• for features i can generate a spec and iterate and generate tasks and iterate.
• There is no magic, each task executes when i want it to
• The directories in which all the specs, tasks, notes etc are placed in is configurable via yaml
• I use agents, commands and skills as needed
• I am currently only supporting claude although it should be easy to support codex, gemini, qwen etc.

I have tried to ringfence the rich thinking that @CasJam has brought to the fore in two slash commands i.e "/bert:plan and /bert:spec" with a third "/bert:task" that touches code. The experiment is now graduated to v0.0.1 and is at bert

edited Oct 18

Anthropic recently announced "claude skills" - this formalizes the pattern of using markdown to instruct the AI engine much as a bash script might - unleashing the power of the underlying filesystem.

So, this has made me revise my approach completely. I have turned agent-os into a "skill" (deprecated) and I have created my own companion skill that instruments both itself and agent-os as needed. I am experimenting with ways in which it can be made more granular and have had some promising attempts. Will share them after they are a bit more mature.

As always thoughts and feedback welcome.

edited Oct 12

After digging a bit deeper and spending some time with agent-os 2.0 - I am going in a slightly different direction i.e. creating an add-on to agent-os. This add-on bert (deprecated) is aware of the filesystem schema that agent-os uses and provides task creation and editing capabailities.

This would allow for one to tap the context and spec that resides within agent-os as one works on individual tasks. The idea remains to find ways to break down the set of valuable tasks that agent-os instructs the AI to do into granular re-accesible directives that can be used.

For example adding a git commit after each task, can be accomplished in a few ways either by using the bert add-on to create a task to do it or of course by modifying agent-os itself to be able to do it.

earlier..
The sophisticated approach that @CasJam has taken with agent-os in getting AI engines to take us through the product development cycle shows great promise.

This got me thinking that if we could disassemble agent-os 's approach and understand its grammar and directives we could get more precise control into how we put out prompts. For e.g. we could do things like build spec for just one feature, re-org the directory structure, sharpen the execute task directive, inherit from and override post_flight_checks etc.

I took a stab at it and put some thoughts and code at <deprecated>

Thoughts, pointers to other attempts etc. would be great

[phpmacher]

Hi,

Unfortunately, I don't understand the goal and idea behind your project.

What is it good for and useful for?
What is its intended purpose?

I have no idea.

[ssr1ram]

@phpmacher thanks for your input on the first rev, here is the 2nd rev

[pradeshc]

This from the original post

This would allow for one to tap the context and spec that resides within agent-os as one works on individual tasks. The idea remains to find ways to break down the set of valuable tasks that agent-os instructs the AI to do into granular re-accesible directives that can be used.

For example adding a git commit after each task, can be accomplished in a few ways either by using the bert add-on to create a task to do it or of course by modifying agent-os itself to be able to do it.