Shahzad Ali

Shahzad Ali · Learning notes

From transformer to agents.

The model sits inside a system that can act.

My picture of how a request becomes a sequence of model calls, tool results and decisions.

SHAHZAD ALI / MY NOTES / 06AI, in pictures

Give the model
a way to act.

An agent combines a model with software that runs tools and returns their results.

Your goal

Research this company and prepare a sourced briefing.

Context · completion criteria · permitted actions

01 / Decide

Model

Read the context. Choose the next step.

02 / Act

Tools

SearchRead filesRun code

The application runs the permitted action.

03 / Observe

Result

New evidence goes back into the context.

Repeat until the task is complete, a limit is reached or a person needs to decide.

The software around the model

Instructions · memory · tool access · approvals · progress · limits

Independent work can be delegated to other agents. Their findings return to the coordinator for review.

  1. The request

    Start with an outcome

    A useful request gives the system an outcome, relevant context and a way to recognise completion. The application defines available tools and limits.

    For an acquisition example, that could be a sourced research report with explicit assumptions, gaps and scenarios for human review.

    How a prompt supplies context
    Anthropic

  2. The transformer

    The model processes the available context

    Text becomes tokens and numerical representations. Transformer layers process those representations. The model can produce an answer or request an available tool.

    A tool call is structured output that surrounding software handles. During ordinary inference, the trained parameters remain fixed.

    Inside the transformer
    Brown and colleaguesOpenAI

  3. The agent loop

    The harness keeps the process moving

    The surrounding software passes instructions and context to the model, executes permitted tools, returns observations and records progress. The loop can continue, finish or pause.

    The model proposes actions within the system it has been given. An application also needs stopping conditions, error handling and clear boundaries.

    The agent loop
    OpenAI

  4. State and retrieval

    Memory is information the system can return to

    Conversation state, files and retrieved passages can carry useful information between steps. Each later call receives the context the application chooses to supply.

    Stored material may be incomplete, outdated or wrong. Finding it again does not establish that it is true.

    How retrieval works
    OpenAIAnthropic

  5. Parallel work

    A team helps when the tasks can separate

    A coordinating agent can send independent questions to specialist agents and combine their findings. Shared dependencies and files need coordination.

    Every agent uses context and computation. Extra reviewers can also repeat an error, so evidence checks and clear task boundaries matter.

    Agents and their tools
    Anthropic

  6. Review and permissions

    Actions need the right authority

    Tools can read information or change external systems. The application can require approval before a consequential action and preserve state while it waits.

    A report, a draft email and a sent email represent different outcomes. The task should make the intended outcome clear.

    Checking the outcome
    OpenAI

Sources behind the technical details

Sources for these notes.

Reviewed . Technical details are linked to their sources. My perspective on how I use AI is personal.

  1. Primary research · First released

    Language Models are Few-Shot Learners ↗

    Brown and colleagues. Describes the 175-billion-parameter model and learning from prompt examples with fixed model parameters.

  2. Living documentation · Living documentation

    Running agents ↗

    OpenAI. Documents model calls, tool execution, handoffs, final outputs and limits in the Agents SDK.

  3. Developer engineering account · First released

    Building effective agents ↗

    Anthropic. Engineering patterns for workflows, autonomous loops and orchestrator-worker systems.

  4. Developer engineering account · First released

    How we built our multi-agent research system ↗

    Anthropic. Describes parallel research, coordination, evaluation and resource costs in one deployed research system.

  5. Living documentation · Living documentation

    Human-in-the-loop tool approval ↗

    OpenAI. Shows how applications pause tool execution for approval and resume from saved state.

  6. Living documentation · Living documentation

    Reasoning models ↗

    OpenAI. Explains reasoning effort and reasoning-token usage. Available settings depend on the model.

  7. Developer announcement · First released

    GPT-5.6 release ↗

    OpenAI. A dated account of the family, its Ultra orchestration and developer-reported evaluations. Pricing updates appear on the page.

  8. Living documentation · Living documentation

    What are tokens and how to count them? ↗

    OpenAI. Gives rough English conversions and explains why actual counts vary.

← Back to my computing notes