The Biggest Leap Did Not Happen Inside the Model
If AI feels far more useful to you this year than it did two years ago, that impression is accurate. The reason behind it, though, is widely misread. The biggest leap of 2026 did not happen inside the model. It happened around it — in the software that gives a model access to tools, memory, a workspace, and feedback.
We covered the perception side of this in AI Feels Smart Because It Can Act, Not Just Think: AI feels intelligent because we finally see real outcomes, not because its reasoning suddenly doubled. This article takes that idea down to the architecture level and maps where we actually stand today.
Three Phases: Prompt, Context, Harness
Phase 1 — Prompt Engineering
The unit of work was a single interaction. How do you phrase an instruction so the model returns the best possible answer in one pass? This is where few-shot examples, chain-of-thought, and role prompting came from. One question, one answer.
Phase 2 — Context Engineering
The question shifted to what information the model should see at the moment it works. Retrieval-augmented generation, memory stores, automatic summarisation, and layered system prompts became standard practice. A perfect prompt is worthless if the right data never reaches the context window.
Phase 3 — Harness Engineering
The vocabulary of "harness engineering" emerged in early 2026, and attribution for the phrase is contested — some accounts trace it to a February 2026 post by Mitchell Hashimoto, others credit Vivek Trivedy of LangChain. The formula most often quoted is blunt: agent = model + harness.
A harness, also called scaffolding, is the software infrastructure surrounding the model: tool dispatch, memory and state persistence, an isolated sandbox or workspace, context management, and guardrails such as scoped permissions, approval tiers, and monitoring. Because an LLM is stateless and produces only text, the harness is what lets it act across multiple steps and sustain work across sessions.
| Phase | Unit of work | Central question | Typical artefacts |
|---|---|---|---|
| Prompt engineering | One interaction | How do I ask? | Prompt templates, few-shot |
| Context engineering | One context window | What does the model see? | RAG, memory, compaction |
| Harness engineering | The whole environment | How does the system recover when the model is wrong? | Tools, sandbox, progress files, tests |
Harness engineering is positioned as the broader layer: it contains the other two as components.
Chatbot, Copilot, and Autonomous Agent: An Architectural Difference
These three words get used interchangeably in business conversations, yet their cost, risk, and build effort differ enormously.
| Aspect | Chatbot | Copilot | Autonomous agent |
|---|---|---|---|
| Core pattern | Input → output | Suggestions inside a human workflow | Loop: goal → plan → act → evaluate |
| State | Practically none | Limited to a session or editor | Persistent across sessions |
| Tool access | None or minimal | Confined to one application | CLI, APIs, browser, file system |
| Who executes | The human | The human accepts or rejects | The system, under supervision |
| Failure point | A wrong answer | A wrong suggestion, easily dismissed | A wrong action with real consequences |
| Harness required | Minimal | Moderate | Heavy and non-negotiable |
The practical implication: adding a chatbot to a website is an integration project. Building an agent allowed to touch production systems is an engineering project that needs sandboxing, layered permissions, and an audit trail.
Engineering Lessons from Anthropic's Experiment
Anthropic documented an experiment in getting agents to work across many context windows. The finding is sobering: even a frontier model such as Opus 4.5, run repeatedly in a loop, falls short of producing a production-quality web app when given only a high-level prompt like "build a clone of claude.ai".
Two failure patterns showed up. First, the agent tried to do too much at once, ran out of context mid-implementation, and left a half-built, undocumented feature behind. Second, later in a project, a fresh agent would look around, see that progress had been made, and declare the job finished.
The fix was not a bigger model. It was a better environment:
- An initializer agent that lays the foundation: an
init.shscript to start the development server, a progress log file, and an initial git commit. - A structured feature list in JSON — over 200 features in the claude.ai clone example, all marked as failing at the start. JSON was chosen because the model is less likely to overwrite or edit it than Markdown.
- Incremental work: one feature per session, closed out with a descriptive git commit and a progress update.
- End-to-end verification through browser automation rather than unit tests or
curlalone, because agents tend to mark a feature complete when it does not actually work for a user.
Every one of those is a harness design decision, not a model selection decision.
Where Indonesia Stands: 2026 Adoption Data
The "Unlocking Indonesia's AI Potential 2026" study from AWS and Strand Partners — surveying 1,000 business leaders and 1,000 members of the general public, released at AWS Summit Jakarta in August 2026 — gives a reasonably clear picture.
| Indicator | 2026 figure |
|---|---|
| Companies that have adopted AI | 40% (up from 25% the previous year) |
| Adopters now running AI in production | 44% |
| Reporting higher productivity | 75% (up from 68%) |
| Saying AI accelerated innovation | 72% |
| Seeing revenue growth | 62% |
| Citing a digital and AI talent shortage as the biggest barrier | 56% |
That 44% is the number that matters most. It means AI in Indonesia has moved past the pilot stage into live operations. It also means more than half of adopters are still stuck in trials — and the barrier they name most often is not the model. It is talent and organisational readiness.
Why the "Smartest Model" No Longer Decides the Outcome
Three reasons.
Swapping models is cheap. In 2026, moving from one model to another is usually a configuration change. What is expensive and hard to copy is the harness: tool integrations, permission rules, proprietary data, and verification processes.
Research keeps showing environment beats size. Harness-1, an open-source search agent, improved retrieval accuracy chiefly by redesigning the software environment around the model rather than by enlarging the model. Related work on Self-Harness has agents mining their own failures to propose and validate changes to their own harness.
The wrapped component is non-deterministic. A harness is designed to recover gracefully when the model fabricates an action or reports a task as finished when it is not. Birgitta Böckeler of Thoughtworks distinguishes the inner harness shipped by the model's builder — an agent SDK or a coding tool — from the outer harness you assemble yourself: instruction files, Model Context Protocol servers, custom skills. She further separates guides, which steer the agent before it acts, from sensors, which observe the result and let the agent self-correct.
For a business owner, the translation is simple: your AI budget should not be consumed by the most expensive model licence. It should go into designing the workflow, the data, and the safeguards around it.
Our Editorial Roadmap for the Next Month
This topic is too large for one article. Over the next four weeks we will break it into pieces you can act on:
- Week 1 — Anatomy of an agent harness. A component-by-component breakdown: tool dispatch, memory, sandboxing, guardrails.
- Week 2 — Context engineering for business data. Building RAG that stays accurate on internal documents, including Indonesian-language corpora.
- Week 3 — Case study: a first internal agent. From a manual process to a semi-autonomous workflow, with real cost figures.
- Week 4 — Guardrails, token cost, and governance. When an agent should stop and ask a human for approval.
AI Services at katili.dev
We help businesses reach the execution phase without hiring an in-house AI team:
- Integrating AI features into an existing website — conversational assistants, semantic search, automatic summarisation.
- An AI blog generator for steady bilingual content production.
- Designing and building internal agents with guardrails and audit trails.
- Hosting, deployment, and maintenance so your AI system stays alive after launch.
- Architecture consulting: deciding whether you need a chatbot, a copilot, or a genuine agent.
Closing
The 2026 map fits in one sentence: intelligence has moved from the model to the system. The companies that win are not the ones with access to the smartest model, but the ones that design the environment around it most carefully. The AWS data shows Indonesia has already crossed the starting line. The question is no longer whether AI is relevant, but how disciplined you are about building the harness around it.
