Agent and Orchestration Frameworks

What Is an LLM Agent

An agent is an LLM that:

1. Receives a goal
2. Decides which tools to use
3. Executes actions (code, searches, APIs)
4. Observes the results
5. Continues until the goal is complete

Components: LLM + Tools + Memory + Planning + Action

---

Autonomy Taxonomy (L1–L5)

A classification framework for agents, analogous to the SAE levels of autonomous driving. It characterizes a system along two axes: scope of delegated decisions (what the agent decides on its own) and duration of operation without supervision (how long it runs without a human checkpoint). Source: survey by Deli Chen / DeepSeek (see below).

Central insight: the barriers to L5 are not raw capability, but rather (1) persistent accumulation of knowledge across sessions, (2) reliable self-evaluation without human supervision, and (3) grounded scaling of agent architectures without degradation as complexity grows.

The frontier in 2026 operates at L4 (autonomous multi-step execution in limited domains); L5 has not yet been reached by any system. The boundaries between levels are not rigid — the same framework can operate at L3 for routine tasks and drop to L2 in novel situations that require human judgment.

Four architectural patterns

The most powerful systems combine several patterns.

Six fundamental open problems

1. Cognitive loops — the agent repeats failed strategies without noticing.
2. Context window limitations — long sessions (>100k tokens) lose the initial information.
3. Novelty assessment — how to gauge whether an AI result is truly original (semantic similarity does not distinguish novel from obscure).
4. Reproducibility — inference with temperature ≠ 0 gives different results on each run; high sensitivity to the prompt.
5. Safety / dual-use — the capabilities that make the agent valuable also create risk.
6. Cost — solving one SWE-bench case costs ~US$ 5–50 in API calls → economic barrier.

Relevance to the Koder Stack

The Stack operates today at L3–L4: Claude Code/Kode is canonically L3 (checkpoints), and the workflow engine (deterministic orchestration with subagents) directly instantiates the patterns of hierarchical orchestration (supervisor → executors) and multi-agent collaboration (adversarial verify, judge panels). The six problems map to real risks of our harnesses: cognitive loops (loop-until-dry with a counter), context window (compaction), reproducibility (schemas + adversarial verification).

Source: Deli Chen (DeepSeek), "From Copilots to Colleagues: A Survey of Autonomous Research Agents" (2026) — 95+ papers, 17 systems, a matrix of 6 dimensions. PDF. Meta curiosity: ~99% of the article was written by the author's own agent framework (Deli AutoResearch SKILL); credited co-authors: DeepSeek-V4-Pro (text) and GPT-Image2 (figures). Do not confuse it with the homonymous paper arXiv:2508.12752 ("Deep Research: A Survey of Autonomous Research Agents"), which is a different work.

---

Autopilots — always-on agents with governed identity

A category of agent formalized by Microsoft at Build 2026 (06022026) with Microsoft Scout: an always-active agent that operates in the background autonomously, without being reprompted for each task (it schedules cross-timezone meetings, prepares materials, anticipates risks). It is distinguished from the reactive chatbot/copilot by two new axes — persistence (always-on) and identity governance — and not by raw capability.

The architecturally relevant point is not the product, it is the identity model: