Micro-Agents and Micro-Packs
Core Thesis
Forge's micro-agent idea is innovative because it resists the impulse to turn every task into a large, open-ended agent loop. A micro-agent is a bounded local-LLM run for a single judgment task using a fixed micro-pack.
The key idea is: use the smallest agent that can responsibly make the judgment.
Condensed Thought
Some tasks do not need a full autonomous coding agent. They need a focused judgment: inspect one UX rule, classify one finding, evaluate one evidence slice, or apply one discipline lens to one artifact. Forge micro-agents are designed for that kind of bounded work.
A micro-agent uses a micro-pack with a manifest and fixed prompt, assembled context, a local_llm_worker, LCDL chat, model output, logs, and a detection gate. The MVP intentionally avoids vector databases and MCP, keeping the execution path inspectable.
Why It Stands Out
This is a cost-aware and risk-aware agent pattern. Instead of assuming bigger loops are better, Forge creates a path for small, repeatable, local judgment cells. That can reduce token cost, reduce blast radius, improve testability, and make evidence easier to review.
Micro-agents also fit the workcell philosophy. They are bounded workcells that return reviewable output rather than owning delivery state.
Forge Ecosystem Hooks
- Micro-pack contains manifest, prompt, and context schema.
- forgesdlc-kitchensink can assemble context for UX auditor use cases.
- forge-workcells can host the local_llm_worker runner.
- LCDL provides governed model calls.
- AgentRun and WorkcellRequest/Result connect the micro-agent to ForgeRun.
- EvidencePacket can include findings, logs, model output, and detection gate results.
Architecture Implications
Micro-agents need strong boundaries:
- One judgment task per run.
- Fixed prompt and manifest in a versioned micro-pack.
- Explicit context schema.
- No hidden retrieval in the MVP unless deliberately introduced later.
- Output parsing and detection gates.
- Logs and artifacts linked to AgentRun and ForgeRun.
- Clear owner for each pack and runner.
- Benchmarks or fixtures to test whether the micro-agent improves detection quality.
This pattern can scale horizontally: many small judgment cells instead of one giant agent brain.
Blog Post Seed Paragraph
Agentic systems often get larger than the task requires. A single judgment can become a sprawling loop with broad context, tool access, retrieval, and unclear authority. Forge's micro-agent pattern moves in the opposite direction. A micro-agent performs one bounded judgment task using a fixed micro-pack, assembled context, governed LCDL call, and parseable output. It is small enough to inspect, test, and attach as evidence.
Risks And Counterarguments
Micro-agents can create fragmentation if every small task invents its own pack, output shape, and scoring method. Forge should standardize pack structure, output schemas, and evidence attachment while allowing domain-specific prompts.