AHMAD BASHEER · OPERATOR BRIEFING · 2026

THE EVOLUTION OF
AI AGENTS

From prompt engineering to context engineering to harness engineering — the three eras that turned a chatbot into a workforce.

PROMPT ENGINEERING

2020 – 2022

CONTEXT ENGINEERING

2023 – 2024

HARNESS ENGINEERING

2024 → 2026
THE OPERATOR

Who is pitching

AHMAD BASHEER
AI SYSTEMS ARCHITECT · BASHEER.APP

I don't study agents from the outside. I run a production agent fleet every day — coder, deployer, CRM operator, delivery, people ops, expenses — each one wired into live business systems and judged by real outcomes, not demos.

This deck is the map I use to build: three engineering eras, and why the third one is where the leverage lives now.

12+
agent lanes running in production — Slack in, business systems out
20+
live platforms shipped on the basheer.app estate
03
eras of agent engineering shipped end-to-end: prompt, context, harness
24/7
the fleet runs on schedules and queues — no human in the loop required
THE THESIS
Every jump in agent capability was a jump in what we engineered: first the words, then the context, then the machinery around the model.
01PROMPTS
craft the completion
02CONTEXT
engineer what the model sees
03HARNESSES
engineer what acts around it
THE MAP

Three eras, one timeline

Each era ended when its core artifact hit a ceiling — and the industry engineered around it.

PROMPT ENGINEERING

2020 – 2022
  • 2020GPT-3 (175B): few-shot in-context learning — no fine-tuning, just examples
  • 2022Chain-of-Thought: ask the model to reason step by step
  • NOV 22ChatGPT ships — prompting becomes a mass skill
The artifact: a well-written prompt

CONTEXT ENGINEERING

2023 – 2024
  • JUN 23Function calling: the model asks, code acts
  • 2023RAG at scale: vector databases feed the window
  • NOV 24MCP: one open protocol for every tool
The artifact: a grounded context window

HARNESS ENGINEERING

2024 → 2026
  • 2025Coding agents go production: Claude Code, Codex, Cursor
  • 2025Agent protocols mature (ACP / A2A): agents hand work to agents
  • 2026Personal agent runtimes (OpenClaw) run skills on schedule
The artifact: a working harness
01
STAGE 01 · PROMPT ENGINEERING

Origins: the prompt becomes the interface

May 2020 — GPT-3 shows that a 175-billion-parameter model learns from examples placed inside the prompt itself. No retraining, no gradients. The prompt stops being a search box and becomes a programmable interface.

MAY 2020
GPT-3
“Language Models are Few-Shot Learners” — in-context learning goes mainstream
2022
InstructGPT
RLHF aligns raw models to instructions — prompting gets a common dialect
2021–22
Prompt marketplaces
prompts traded as artifacts; “prompt engineer” becomes a job title
INSTRUCTIONS + EXAMPLES
LLM
THE COMPLETION
Source: Brown et al., “Language Models are Few-Shot Learners,” NeurIPS 2020.
01
STAGE 01 · PROMPT ENGINEERING

Core techniques

2020
Zero-shot
state the task, trust the prior — “classify this review as positive or negative.”
2020
Few-shot
show 2–5 examples in the prompt; the model imitates the pattern.
2021
Role prompting
“You are a senior auditor…” — persona steers tone, depth and format.
2022
Chain-of-Thought
“think step by step” — intermediate reasoning lifts math & logic scores dramatically.
2022
Self-consistency & ReAct
sample several reasoning paths and vote; interleave reasoning with tool actions.
Sources: Wei et al. 2022 (CoT); Yao et al. 2022 (ReAct); Wang et al. 2022 (Self-Consistency).
01
STAGE 01 · PROMPT ENGINEERING

Where it hit the wall

Stateless
every call starts from zero — nothing is remembered between requests.
Context ceiling
2–4K tokens in the GPT-3 era; whole knowledge bases simply didn't fit.
No actions
the model could write about your CRM, but it could never touch your CRM.
Brittle
reword the same request and behavior silently changes.
Hallucination
fluent, confident, unverified — and nothing in the prompt grounds it in fact.
The fix wasn't better words. It was better surroundings — what the model sees, remembers, and can do.
02
STAGE 02 · CONTEXT ENGINEERING · 2023 – 2024

Why context is everything

An LLM is a stateless function: output = f(weights, context). You can't fine-tune your way out of every gap — but you fully control what goes in the window. Context engineering is designing that input on purpose: the right facts, the right tools, the right history, at the right moment.

Retrieval beats retraining: swap knowledge by swapping context
The window is working memory — manage it like RAM, not a dump
“Lost in the middle”: position in the window matters as much as content
CONTEXT WINDOW · THOUSAND TOKENS
2
GPT-3
2020
4
GPT-3.5
2022
32
GPT-4
2023
200
Claude 2.1
2023
1,000
Gemini 1.5
2024
Sources: OpenAI, Anthropic, Google model cards, 2020–2024.
02
STAGE 02 · CONTEXT ENGINEERING

RAG: retrieval before reasoning

1 · QUERY
user asks; question is embedded into a vector
2 · RETRIEVE
vector search over the knowledge base returns top-k chunks
3 · AUGMENT
chunks packed into the prompt beside the question
4 · GENERATE
the model answers from supplied evidence, with citations
WHY IT WON
Grounding: answers cite retrieved evidence — hallucination drops where it matters
Freshness: swap the index, not the model — knowledge updates in minutes, not retraining cycles
Access control: filter retrieval per user, so agents only see what they're allowed to see
Stack in the wild: pgvector, Pinecone, Weaviate, Qdrant. Source: Lewis et al., “Retrieval-Augmented Generation,” 2020.
02
STAGE 02 · CONTEXT ENGINEERING

Memory & tools: from text to action

MEMORY
Short-term
the context window itself — the working set for this task
Long-term
embeddings + vector recall across sessions; the agent remembers you
Episodic
structured session logs that can be replayed and reflected on (MemGPT, 2023)
TOOL USE
Function calling
Jun 2023 — the model outputs a structured request; your code executes it
Plugins
2023 — browsers, code runners, calendars; actions become first-class
MCP
Nov 2024 — one open protocol: any tool, any model, any vendor
THE LOOP THAT CHANGED EVERYTHING
model asks → harness executes → result returns to the window → model decides again
03
STAGE 03 · HARNESS ENGINEERING · 2024 → 2026

The harness: engineering around the model

A harness is everything that isn't the model: the loop, the tools, the permissions, the file system, the memory, the skills. The model proposes; the harness disposes.

GOAL
outcome in plain language
PLAN
decompose into steps
ACT
run tools in a sandbox
OBSERVE
read results, re-plan
↺ OBSERVE loops back to PLAN — until done, or until a budget stops it
WHAT THE HARNESS OWNS
Executor loop — until-done iteration with budgets and stop conditions
Sandbox & permissions — what the agent may touch, what needs a human yes
Sub-agents — specialists spawned per task, results merged
Skills & playbooks — tested procedures the agent loads on demand
Context compaction — summarize and prune so long runs keep moving
Observability — traces, evals and audit trails on every action
STAGE 03 · HARNESS ENGINEERING

Framework evolution, 2023 → 2026

MAR 2023

AutoGPT

the first consumer autonomous agent — give it a goal, it loops. 100K+ stars within months.
APR 2023

BabyAGI

task queues + re-prioritization — the plan/act/observe loop made legible.
2022 → 24

LangChain → LangGraph

from prompt plumbing to graph-based orchestration; agents become engineering artifacts.
2024 – 25

Coding agents

Cursor, Claude Code, Codex: harnesses with file systems, terminals and permission gates.
2025 – 26

OpenClaw

personal agent runtime: lives in your channels, runs skills on schedule, holds memory.
THE GRAVITY SHIFT
2023's frameworks arranged prompts into chains. 2025's harnesses own state, tools, permissions and time. The framework became an operating loop — and the loop became the product.
Star count: GitHub trending data, 2023 (approx.).
STAGE 03 · HARNESS ENGINEERING

Representative products

Three products, three answers to “what should a harness be?”

2023

AutoGPT

THE PROOF OF DEMAND
  • Autonomous goal loop: plan, act, observe, repeat until done
  • First agent most of the world ever saw — 100K+ stars in months
  • Also the proof of the reliability ceiling: loops drift, costs run away
2025 →

ACPAgent

AGENTS THAT TALK TO AGENTS
  • Built on ACP — the Agent Communication Protocol
  • Standard interface: agents discover, message and hand off work across vendors
  • The harness stops being one app and becomes a network of specialists
2026

OpenClaw

THE HARNESS AS A PRODUCT
  • Open-source personal agent runtime — messaging-first
  • Skills loaded on demand, cron schedules, durable memory, multi-channel
  • What I run daily: it's the shape of stage-three computing
SIDE BY SIDE

Three eras, one comparison

ERAUNIT OF WORKKEY ARTIFACTFAILURE MODELEVERAGE
01 · PROMPTone completionthe prompt itselfsilent drift on rewordingwording craft
02 · CONTEXTa grounded taskRAG + memory + tools in the windowstale or wrong contextretrieval & curation
03 · HARNESSan outcome, end to endloop + tools + permissions + skillsrunaway or unsupervised actionssupervision & evals

Read the table as a stack, not a history: 2026 harnesses still contain 2022 prompts and 2023 retrieval. Nothing was replaced — each era became a layer.

TRAJECTORY

Where the leverage moves next

01
Skills become the unit of IP
tested playbooks a harness loads on demand — my 90-skill catalog is exactly this.
02
Fleets replace assistants
orchestrator + specialist agents, each scoped, each observable, handing off over protocols like ACP.
03
Agents on the payroll model
metered, scheduled, accountable — work measured in outcomes, not tokens.
04
The moat is the harness
models commoditize; context strategy, tool surface and supervision don't.
The prompt was the interface. The context is the engine. The harness is the product.
THE ASK

I build at Stage 03 today.

While the market debates prompts, I operate a twelve-lane agent fleet wired into a live enterprise — CRM replacement, proposal engine, delivery pipeline, people ops — running on schedules, not good intentions.

If your organization wants that layer built, that's the conversation.

AHMAD BASHEER
AI SYSTEMS ARCHITECT · BASHEER.APP
THIS DECK LIVES AT PITCHING.BASHEER.APP
AHMAD BASHEER — AI AGENTS
← → NAVIGATE
01 / 16