Scripts that click buttons and fill forms. Brittle, deterministic, zero reasoning ability.
ELIZA (1966) through scripted IVR bots. If-then logic with no language understanding whatsoever.
Knowledge engineers encoding domain expertise. Powerful in narrow lanes — useless outside them.
OpenAI and Anthropic bake structured tool use directly into the model API. No hacks required.
Opinionated scaffolding for building agents with memory, tools, and multi-agent coordination.
The field pivots from demos to production. Evaluation, cost, and error recovery dominate the conversation.
Vector databases, RAG pipelines, agent observability platforms, prompt management — the stack is forming.
Claude Code, GitHub Copilot Workspace, Devin — agents that own entire feature branches end-to-end.
Models that see screens and use any software like a human. Anthropic Computer Use is in production.
Agents running scheduled tasks, managing workflows, filing reports — without per-step human supervision.
Safety-first, long context, best-in-class coding agents. Claude Code. Computer Use. Pioneer in tool use design.
Broadest ecosystem, Agents SDK, widest plugin support, strongest consumer brand recognition.
Deepest Workspace integration, 1M+ token context, multi-modal leader, strong search grounding.
Enterprise Office and Teams integration. Dominates corporate IT deployments at massive scale.
Each step in an agent chain: ~90% success. After 10 steps: 35% overall. Errors cascade and amplify — fast.
Token costs multiply rapidly in agentic loops. Long-horizon tasks can exhaust budgets before completing.
How do you score an agent that ran 40 steps? Human eval is slow. Automated eval is imperfect. Both are needed.
When can an agent take an irreversible action? How does a human stay in the loop without killing throughput?
What do I need to do? What information do I have? What tool should I call?
Call a tool. Write to memory. Send a message. Execute code. Make an API call.
Parse the result. Did it succeed? What data came back? Any errors to handle?
Update working memory. Revise the plan. Loop back to Reason — or finish.
Reads the ticket (goal), decides which station to call (tool), checks the pass (observes result), updates the board (memory), calls the next station (loops).
Grill, sauté, garde manger — each specialist does one thing fast and well. The expeditor coordinates; never cooks alone.
Ticket rail shows what's been called, what's pending, what's fired and waiting. Lost context equals lost ticket equals chaos.
Every round of service is a ReAct loop. Multiple loops per cover. Service ends when all tickets are cleared.
Model outputs structured JSON describing a function call. Runtime executes it and returns results into model context.
The model never actually runs code. It requests execution. Your infrastructure decides what it's allowed to do.
Everything currently in the model's attention. Fast, ephemeral, limited. Lost when the session ends.
Vector databases, files, SQL tables. Persists across sessions. Retrieved via semantic or exact search.
A log of what the agent did, what happened, what worked. Improves future decisions, avoids past mistakes.
| Dimension | Chatbot | Agent |
|---|---|---|
| Execution | Single response | Multi-step autonomous loops |
| Memory | Session context only | Persistent external memory |
| Tools | None or very limited | Many — code, web, files, APIs |
| Autonomy | Waits for each prompt | Runs until goal is complete |
| Planning | No | Yes — decomposes and sequences |
| Error handling | Reports error, stops | Retries, replans, recovers |
Open ChatGPT → Click "Explore GPTs" → Click "Create"
Describe your agent's purpose in plain English — the builder configures it conversationally
Add capabilities: web browsing, code interpreter, image generation, custom tool Actions
Upload knowledge files, connect OpenAPI schema Actions, configure and publish
Quick prototypes, non-technical builders, internal tools with simple logic, customer-facing GPT wrappers
No persistent memory by default, limited customization, locked to OpenAI infrastructure, cannot run background tasks
Full control, custom memory, any infrastructure, production-grade, no vendor lock-in
Custom workflows, sensitive data, complex multi-step logic, existing system integration
Python or TypeScript, API key management, your own hosting, strong prompt engineering
Not a chatbot with tools bolted on. Designed from the ground up for persistent, goal-directed execution with memory.
Connect external services, run background tasks, maintain memory across sessions without custom infrastructure.
Delivers agentic capability to users who can't or don't want to write Python. The GUI layer over agent infrastructure.
Platforms that bake in agentic thinking now will be years ahead of those trying to retrofit it later.
Best for RAG pipelines, complex chains, huge ecosystem of integrations. Use when you need flexibility and can afford config time.
Best for defining agent roles and letting them collaborate on a task. Opinionated, readable, great for team-based workflows.
Best when agents need to debate, critique, and refine each other's work. Microsoft-backed. Strong for research and analysis.
This is the orchestrator-worker pattern running in production. Claude Code orchestrates. Bash, file tools, compilers are workers. You describe the outcome; it engineers the path.
Cross-session memory stored externally. Garde Manger remembers context between conversations — not a chatbot reset.
File access, code execution, web browsing — a genuine tool-using agent, not a text generator with a friendly face.
Given a goal, it works toward completion across multiple steps without requiring a human prompt for each action.
Claude Code acts as orchestrator — planning, delegating to tool-workers (bash, file system, compiler), aggregating results, iterating. Textbook multi-agent architecture in production.
Using multiple model instances — or models from different providers — to review, critique, and refine each other's outputs. The evaluator-optimizer pattern in action.
Describe desired outcome in natural language → Agent decomposes → Agent executes → Agent reports → Human reviews output, not process
You are operating at the goal level, not the task level. This is leverage. One well-formed description replaces hours of manual execution by a human.
A full specification for a personal AI agent with persistent identity, memory, proactive behavior, and multi-modal tool use.
Not a chatbot spec. A companion agent spec — context-aware, relationship-maintaining, autonomously improving over time.
This spec describes what most AI labs are working toward for their consumer products. You wrote it as an independent developer.
One agent decomposes goals and routes sub-tasks to specialist agents or tools. Scales to complexity. Already in daily use with Claude Code.
Agents that read and write persistent external memory. Context accumulates over time. Garde Manger is this pattern, fully realized.
One agent generates, another evaluates, the first refines. Produces higher quality outputs than any single-pass generation. Running with Gemini.
Background agents that activate on schedules, webhooks, or data thresholds. Proactive rather than reactive. Heartbeat is this pattern.