Ology

Beyond the Chatbot: Engineering Agentic Systems

AI agents and agentic architectures: components, planning, tool use, memory, coordination, execution, engineering, evaluation, deployment, limitations, and open research questions.

5 sessions · 15 readings · 11 views

Curated by Rajeev Chhajer

The Control Surfacesession 1Tools, Context, and the Shape of Thoughtsession 2Memory Is a Runtimesession 3Parallel Minds, Shared Limitssession 4Trust at the Boundarysession 5

Opening

Agentic systems have moved from demonstrations of tool-calling models toward software that can search, write code, operate computers, query internal data, delegate subtasks, and continue working across long-running sessions. The important shift is not simply that models have become better at reasoning; it is that engineers are building increasingly elaborate loops around them—loops that manage context, tools, state, permissions, evaluation, and failure recovery.

Recent technical work has also made the boundaries of the system less obvious. Tool definitions can overwhelm a context window. Memory may mean retrieval, institutional knowledge, durable artifacts, or simply a carefully designed harness. Multi-agent systems can improve breadth while multiplying cost and coordination failures. And the latest production guidance increasingly treats the agent runtime—not the prompt—as the place where reliability and security must be enforced.

The resulting design space contains real disagreements. Anthropic argues for simple composable patterns and selective autonomy; OpenAI’s platforms increasingly package orchestration, environments, tracing, and delegation as reusable infrastructure. Other work shows that code execution, parallel agents, and long-horizon harnesses can unlock new capabilities, but also expose weaknesses in planning, evaluation, and security.

This seminar follows the central engineering problem: how much control should belong to the model, and how much must remain in the surrounding system?

Five questions worth arguing about

  1. 1When should an application let a model control execution rather than constrain it with an explicit workflow?
  2. 2Can better tools and context engineering substitute for better planning, or merely make failures more expensive?
  3. 3Is agent memory primarily retrieval, durable state, or a disciplined way of handing work between sessions?
  4. 4Do multi-agent systems create new problem-solving capacity, or mostly purchase performance with tokens and complexity?
  5. 5Can evaluation and security constraints make autonomous agents trustworthy without making them too constrained to be useful?

The Sessions

1Session 1start here

The Control Surface

The problem

The first design decision is not which framework to adopt, but where to place control. Anthropic distinguishes workflows, whose paths are specified in code, from agents, whose models decide how to proceed. Its recommendation is deliberately conservative: begin with the simplest system that can work, and introduce autonomy only when the task’s structure genuinely requires it.

Hugging Face’s account makes the same boundary concrete by describing agency as a spectrum: a model may choose a route, select a tool, or control an entire loop. Its ReAct examples show why memory, parsers, error handling, and retry logic appear as soon as the model controls program flow. OpenAI’s platform launch complicates the picture from another direction: APIs now package multi-turn tool use, computer interaction, orchestration, and observability as standard building blocks.

Together, these readings ask whether “agent” names a fundamentally new architecture or a familiar program with a model-controlled control surface. That distinction matters because every increase in model authority trades predictability for flexibility, and the next session examines the interfaces on which that trade depends.

Readings

  1. Building Effective AI AgentsAnthropic
    Open ↗
  2. Open-Source LLMs as LangChain AgentsAymeric Roucher
    Open ↗
  3. New Tools for Building AgentsOpenAI
    Open ↗
2Session 2Tools, Context, and the Shape of Thought3 readings3Session 3Memory Is a Runtime3 readings4Session 4Parallel Minds, Shared Limits3 readings5Session 5Trust at the Boundary3 readings

Next Up