tackit - keeping agents on task

After working on multiple large-scale projects with AI coding agents, I built a utility called tackit to help keep agents on task & improve context window efficiency. Here are the problems I'm trying to solve,

• Scattered records: agents constantly create new files in different formats, scattered everywhere, to keep track of tasks which often get buried & lost.
• Context erosion: every prompt re-reads lengthy planning docs, todo files, chat histories, triggering frequent auto-compaction and quality degradation.
• Dependency hell: dependency tracking is almost non-existent, relying on code-execution errors for dependency discovery, burying any where test coverage falls short.

I've been dog-fooding tackit for a week and fixed a ton of logical & efficiency issues, hopefully this little tool can be of help to others.

tackit keeps agents on track

tackit is a deterministic app that helps keep agents on track with task management (in db with strict schema and pydantic validation), strongly nudges agents to resolve dependency cascades (strongly-bounded) upon edits with close-gates (can't close a task without resolving dependency cascades), minimal context bloat by surfacing & retrieving exactly the specs needed to complete tasks, and diff-based editing to minimize context & inference bloat.

It also comes with a wide range of skill, mcp, and agent settings (e.g., claude.md) based nudges to improve coding agents' behaviors. Failing loud / ship-on-pain clauses help to revert agents' natural inclination to frequently ignore failures, assume degradation is acceptable, and silently moving forward. A fold-back behavior that strongly suggests folding learnings back into specific task's design & implementation specs post-implementation, as the process of writing code often leads to new discoveries.

There are a ton more nuanced features & capabilities but these are the highlights.

opinionated spec<>task coupling

tackit is natively aware of the difference between design / schema specifications, vs. production (coding) tasks, vs. meta (discussion, notes, comments) tasks. Specifically tackit is designed to tightly couple specification <> production tasks.

After using coding agents for a while, I got into the habit of keeping a set of design & schema specs that are constantly synced with implementation. This wasn't usually the case pre-ai, and the design spec written at the beginning of projects tend to to stale pretty quickly, relying on in-code notes & comments as reminders. With ai agents, I found if there wasn't a persistent, clearly laid out spec that is constantly updated, the agent drifts extremely fast and becomes extremely slow as it needs to ingest massive amounts of codebase & run tests to figure out what is happening.

Hence one of tackit's primary goals is to keep specifications constantly in sync with implementation, including all dependencies.

getting started

There are two parts to the installation, one for the human, one for the agent.

Step 1: Humans just run pypi to install the package,

pip install tackit

Step 2: After the package is installed, go into the coding agent and ask it to wire up the mcp & skills.

I've installed the `tackit` Python package. Run `tackit setup` and carry out the steps it prints to register it (MCP server + skill) for this project.

That's it, let me know what you think!