Agentic foundations

Agentic foundations are the setup an agent needs before a model or a framework is chosen: files that serve as memory across runs, and scripts for what repeats.¹ In this view the model and the framework come after that setup, and a chat window is not the memory layer.¹

Overview#

Scoped project context packs and stable profile facts are separate mechanisms, and both differ from context pasted into a single chat.² Facts that a whole team relies on are kept apart from both (see Shared brain). An agent suits work whose steps are unpredictable; where the steps are known in advance, a workflow is preferred.³

Mechanism#

1. Context is kept in files under version control.¹
2. Anything done more than twice is scripted.¹
3. Scoped project context and profile facts are loaded at the start of work,² and team facts are kept in a shared store.
4. The observe, reason, and act cycle is implemented directly before a framework is adopted.³
5. When an answer must be grounded, retrieval is debugged first and the source is cited.³

Applications#

The approach applies when agent work fails because the files or the repeated steps were never set up.

Limitations#

A chat window does not serve as the memory layer, and work done by an agent still needs a check.¹ An agent is not the right tool for steps that already form a workflow.³

Worked example#

The source post sets out an order of setup.¹

1. The project context lives in a private git repository that serves as the memory layer: each agent pulls, works, and merges back, so context that would die in a chat window lives in the repo.
2. Anything done more than twice, such as setup steps and boilerplate, is scripted.
3. The agent is asked for the config and the errors, and what it hands back is double-checked.

See also#

• Multi-agent teams – when to use agents and when to use workflows
• Shared brain – a write-gated team wiki, as distinct from runtime persistence
• Skillify authoring – turning repeated work into a skill
• Verification and stop conditions – checks on agent output
• Harness runtime – loop, tools, permissions, compaction, and hooks

References#

Footnotes#

1. https://x.com/sudoingx/status/2055950007036162207 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

2. https://x.com/eng_khairallah1/status/2056670039101841411 ↩ ↩²

3. https://x.com/deronin_/status/2033587293064204349 ↩ ↩² ↩³ ↩⁴