Agent Context Compression

Generic instructions for reducing token usage across an agent system without reducing correctness, safety, or task completion quality.

Product Context

Context growth is both a cost problem and a reliability problem. This reference architecture treats compression as a product control: preserve decisions, constraints, evidence, and safety boundaries while removing repeated instructions and low-value retrieval.

Maintained by Jason Agentic.

Purpose

Use this when you want an agent to:

• be concise by default
• minimize total system-wide token usage
• avoid redundant retrieval and delegation
• preserve correctness, safety, and completion quality

What this is

This repo contains a reusable instruction file that can be dropped into many agent harnesses or used as a system/developer prompt fragment.

It is intentionally:

• harness-agnostic
• provider-agnostic
• identity-neutral
• focused on token efficiency without unsafe shortcuts

Files

• AGENT_CONTEXT_COMPRESSION.md — the reusable instruction set
• LICENSE — MIT license
• .gitignore — basic exclusions

When to use it

Use these instructions when:

• the agent runs often and token cost matters
• multiple agents may duplicate work
• prompts tend to bloat over time
• you want stricter discipline around delegation, retrieval, and verbosity

Do not use it to justify guessing, hiding uncertainty, or skipping critical warnings.

How to use it

Option 1: paste it into a system prompt.

Option 2: include it as a developer instruction file.

Option 3: merge selected sections into an existing orchestration prompt.

Quick install

1. Copy AGENT_CONTEXT_COMPRESSION.md into your prompt or instruction layer.
2. Place it after role/identity instructions and before task-specific rules.
3. Keep local tool, safety, and product constraints in separate files or sections.
4. Test with one simple task and one multi-step task to confirm the shorter style still preserves correctness.

Copy-paste example

Your operating priority is to minimize total token usage across the whole agent system without reducing correctness, safety, or task completion quality.

[Paste AGENT_CONTEXT_COMPRESSION.md here]

Compatibility

Works best with:

• system prompts
• developer prompts
• agent policy files
• orchestrator instruction blocks

It is not tied to any specific model provider, framework, agent runtime, or tool harness.

Adaptation guidance

You can safely customize:

• default verbosity
• delegation thresholds
• when to ask clarifying questions
• output formatting preferences

Keep these parts intact:

• correctness over brevity
• no unsafe guessing
• no hiding uncertainty
• no omission of critical constraints or failure conditions

Safety notes

• Do not use token efficiency as a reason to skip verification where verification is required.
• Do not suppress warnings that materially affect decisions.
• Do not remove context that changes correctness.
• If brevity and correctness conflict, preserve correctness.

Suggested use cases

• orchestrator agents
• coding agents
• task-routing agents
• tool-using assistants
• multi-agent systems with shared budgets

License

MIT