Engineering Briefing · September 2026

AI in Software
Development Where it stands, and how we use it

A working brief for the engineering team: the current model and tool landscape, the practices that separate teams who ship from teams who thrash, and the risks we have to engineer around.

Prepared by Tafareed Data current as of 14 September 2026
ContentsOverview

What this briefing covers

Six parts, building from what the technology actually is, through how to work with it, to the platforms we can use.

01

Understanding intelligence

What human and artificial intelligence each are, how they differ, and why capability is jagged.

02

Where the field is

Almost everyone uses it, fewer than half trust it, and tools have moved from autocomplete to agents.

03

Models and tools

Models change fast, open-weight options matter, and the harnesses around them are what you actually use.

04

Prompt engineering

The six building blocks, the six common mistakes, and how to refine until a prompt is reliable.

05

Working with agents

Automation versus agentic workflow versus agent, which to build, and context engineering.

06

Automation platforms

Zapier, Make and n8n: what each is and why they matter to us.

2
01

Understanding
intelligence

Before we use it well, we should be precise about what it is — and what it is not. Most bad decisions about AI start with a category error here.

Understanding intelligence01 · Definition

Human intelligence

The full spectrum of cognitive, emotional, social and physical capacities that let a person understand the world, learn from very few examples, and make value-based decisions in ambiguous situations.

Capacity

Reasoning

Logic, abstraction, and cause-and-effect thinking that generalises to situations never seen before.

Capacity

Emotion

Recognising and managing feelings, in yourself and in the person across the table.

Capacity

Creativity

Original thought and genuine invention, not only recombination of what already exists.

Capacity

Judgement

Deciding what should be done, not merely what can be done. This is the one that carries accountability.

A child touches a hot stove once and avoids fire for life. One example, permanent generalisation, no dataset.
You see a colleague's face and know something is wrong before a word is said — an emotional read and a judgement call in the same instant.
4
Understanding intelligence01 · Definition

Artificial intelligence

Computer systems that perform tasks once thought to require human thinking — by learning statistical patterns from very large datasets and using them to predict the next best output.

Property

Pattern-based

Learns correlations from billions of examples. It is extraordinarily good at what resembles what it has seen.

Property

Speed and scale

Runs continuously, consistently, across many tasks at once, without fatigue or mood.

Property

Narrow

Excellent at the shape of task it was trained for, and unreliable a short distance outside it.

Property

Dependent

Needs human framing, evaluation and correction. It has no independent way to know whether it succeeded.

A spam filter saw millions of junk messages, learned the patterns, and now sorts new mail in milliseconds — far beyond any person's capacity.
The same class of system drafts an email instantly and fluently, and may confidently invent a name, a figure or a citation that does not exist.
5
Understanding intelligence01 · Side by side

Two different kinds of mind

Not better and worse. Different in kind — which is precisely why the division of labour between them is a design decision worth making deliberately.

DimensionHuman intelligenceArtificial intelligence
Learns fromLived experience and a handful of examplesEnormous datasets and millions of examples
StrengthDepth and meaning — knowing what mattersSpeed and scale — knowing what, and how fast
CreativityOriginal intent and directionRecombining existing patterns fluently
EthicsMoral agency and accountabilityNo moral agency — it optimises against data
NatureGeneral, contextual, and slow to scaleNarrow, literal, and scales instantly
Failure modeTired, biased, inconsistent — and usually knows itConfident, fluent, and wrong — and cannot tell
The division that follows from this table: AI handles the what and the how fast. People handle the why and what matters. Every failure in section 07 is a case of that line being crossed.
6
Understanding intelligence01 · The key insight

"Jagged" intelligence

Capability is not a smooth line from simple to hard. Even the strongest 2026 frontier models solve doctoral-level problems one moment and fail at something a child handles the next — and nothing in the output signals which you just got.

Why it happens

Fluency feels like understanding, but the model has no model of truth — only of language. It predicts what a correct-sounding answer looks like. When the pattern is right, the answer is right. When the pattern is right and the fact is wrong, the answer still sounds right.

What it means for us

You cannot delegate on the basis of difficulty. "This task is easy, so it will be fine" is not sound reasoning about a jagged system. You delegate on the basis of whether the output can be checked — which is the whole argument of section 05.

It writes complex, working code across several files — and miscounts the letters in a short word.
Asked for a source, it produces a real-looking book and author that have never existed.

The sentence to carry through the rest of this deck

AI is a remarkably capable tool — not a colleague with judgement. It amplifies your thinking; it does not take over your responsibility. Keep a person in the loop wherever the answer has consequences.

7
02

Where we are

Adoption stopped being the interesting question about eighteen months ago. Effectiveness is the interesting question now.

Where we are02 · Adoption

Everyone uses it. Fewer than half trust it.

These two facts sit side by side in every major 2025–2026 survey, and they explain why "just give the team Copilot" is not a strategy.

90%
of technology professionals use AI at work
DORA, 2025 State of AI-assisted Software Development
85%
of developers regularly use AI tools for coding; 62% rely on at least one assistant, agent or AI editor
Stack Overflow Developer Survey 2025, n≈49,000
46%
actively distrust the accuracy of AI-generated code
Stack Overflow Developer Survey 2025
55%
agent-mode adoption, projected above 70% by the end of 2026
Industry tracking, 2026
The most common complaint is not obviously bad output. It is code that looks correct and contains a subtle error — which is exactly the failure mode that review processes built for human mistakes are worst at catching.
9
Where we are02 · Tool generations

Three generations of tooling — all still in the building

Most teams are running a mix. Knowing which generation a tool belongs to tells you how much supervision it needs.

2021 – 2023

Autocomplete

Inline next-line suggestions. You stay in the driver's seat the entire time. Low risk, low ceiling.

Copilot inline, Tabnine, Codestral FIM

2024 – 2025

Chat & IDE assistants

Ask, paste, apply a diff. The model sees files you point it at. You still assemble the context by hand.

Cursor, Cline, Copilot Chat, Continue

2026 — now

Coding agents

Plan, edit across files, run the test suite, read the failure, fix it, repeat. Sessions run for hours. Fleets run in parallel.

Claude Code, Codex, Antigravity CLI, Muse Code, Devin

The consequence for us

When the agent can run the code, the bottleneck moves off "can it write the function" and onto what the agent is allowed to see, run and merge. That is an engineering problem, not a prompting problem — and it is the subject of section 03.

10
03

Models & tools

A snapshot that will be partly wrong in ninety days. Learn the shape of the landscape, not the leaderboard.

Models & tools03 · Release pace

2026 so far: a frontier release every ten days

Selected coding-relevant releases. The point is not to track all of them — it is to notice that any tool decision you make has a ninety-day shelf life.

Feb – Mar
  • Gemini 3.1 ProGoogle DeepMind
  • Claude Opus 4.6first 1M-context Opus
  • GPT-5.3-Codexagentic coding
  • GLM-5745B MoE, MIT
  • GPT-5.4native computer use
  • Nemotron 3 SuperNVIDIA, open weights
Apr – May
  • GPT-5.5OpenAI
  • DeepSeek V4 Pro / FlashMIT, 1M context
  • Kimi K2.6open weights, 1T MoE
  • Claude Opus 4.7 → 4.8parallel subagents
  • Gemini 3.5 Flashpowers Antigravity 2.0
  • Cursor Composer 2.5
Jun – Jul
  • Claude Sonnet 5most agentic Sonnet
  • GLM-5.2 · MiniMax M3open weights
  • Claude Opus 5new Max default
  • GPT-5.6 Sol / Terra / Lunathree capability tiers
  • Kimi K3first open 3T-class
  • InklingThinking Machines
Aug – Sep
  • Qwen3.8-Max / 27BApache 2.0 sibling
  • Grok 4.6500K context
  • GLM-5.3 / 5.3-FlashZ.ai
  • Muse Spark 1.2 → 1.3Meta
  • Claude Fable / Mythos 5.11 Sept
  • Gemini 3.8 Flash · GPT-6 Astra2–3 Sept
Practical rule: standardise on a tier (frontier / balanced / cheap-and-fast) and a harness, not on a specific model name. Swapping the model behind a good harness is a one-line change.
12
Models & tools03 · Open weights

The open-weight tier is now a serious cost lever

For internal tooling, bulk refactors, test generation and anything that runs in a loop, the gap to frontier is small and the price gap is one or two orders of magnitude.

Z.ai · Aug 2026

GLM-5.3 & GLM-5.3-Flash

5.3 targets software engineering, autonomous agents and security, with a claimed 50% coding improvement over 5.2. The Flash variant is multimodal at roughly $0.075 / $0.25 per million. Coding plans start around $18/month. Note 5.3 moved off MIT to a bespoke licence.

Moonshot · Jul 2026

Kimi K3

First open 3T-class model — 2.8T total parameters, native vision, 1M context. Weights on Hugging Face under a custom licence with a commercial-agreement threshold for large model-as-a-service businesses.

DeepSeek · Aug 2026

DeepSeek V4 Pro

Generally available across app, web and API with three reasoning-effort levels and OpenAI Responses API compatibility. 1M context, output to 384K. Off-peak usage bills at half the peak rate — relevant for overnight batch work.

Alibaba · Aug 2026

Qwen3.8-27B

Dense 27B under Apache 2.0, text/image/video in, 262K context extensible to 1M. The realistic option if we ever need to run something on our own hardware for data-residency reasons.

Licence discipline matters more than it used to. "Open weights" in 2026 frequently means a custom licence with revenue thresholds and review clauses, not MIT or Apache. Read the licence before it goes anywhere near client work.
13
Models & tools03 · Harnesses

The harnesses worth knowing by name

The harness is the part that holds context, runs tools and manages subagents. On the current evidence it matters roughly as much as the model.

Terminal agents

Claude Code

Terminal, desktop and web. Subagents, agent teams, skills, hooks and a plugin marketplace. Highest satisfaction of any tool in the JetBrains January 2026 pulse — 91% CSAT on 18% adoption.

Terminal agents

OpenAI Codex

App, CLI, IDE and web. Adopted the SKILL.md format as an open standard, so skill packs now work across both ecosystems.

Terminal agents

Antigravity CLI · Muse Code · Grok Build

Antigravity CLI replaced Gemini CLI at I/O 2026. Meta's Muse Code shipped in beta in August; xAI open-sourced Grok Build in July.

AI IDEs

Cursor 3

Full-screen Agents Window for parallel multi-agent execution across local, cloud, SSH and worktrees, plus a separate CLI with cloud handoff.

AI IDEs

Windsurf 2.0 · Antigravity 2.0 · Kiro

Agent command centres and spec-driven workflows built into the editor. Kiro in particular is built around specs rather than chat.

Open source

OpenCode · Cline · Roo · Aider · Crush

Model-agnostic, self-hostable, no lock-in. OpenCode reported 150K+ stars and 6.5M+ monthly actives by April 2026.

Our recommendation

Pick one primary harness for the team and one open-source fallback. Shared muscle memory, shared config files and shared skill packs are worth more than each developer optimising their own setup.

14
04

Prompt
engineering

The craft of telling a model exactly what you need — clearly, with the right context, in a way that can be repeated and tested.

Prompt engineering04 · The mindset shift

It is a workflow now, not a bag of tricks

In 2026 the job is not finding magic words. It is assembling the right context — role, task, constraints, examples and output format — so the model can reliably do the thing you need, again tomorrow, for someone else.

Principle

Cleverness in wording is over

Cleverness in context assembly is the craft that replaced it. What the model can see matters far more than how you phrase the ask.

Principle

Treat prompts like code

Version them, test them, review the changes. A prompt that matters to the business belongs in the repository, not in someone's chat history.

Principle

Context beats phrasing

Relevant context yields far larger gains than any rewording ever will. If output is poor, ask what the model could not see — not how to say it more nicely.

"Write me a post" → "Write an Instagram post for a new café: friendly tone, three lines, one emoji, and a clear call to visit."
"Fix this code" → "This Python raises a TypeError on line 10; it should sum a list. Fix it and explain the cause briefly."
16
Prompt engineering04 · Anatomy

The six building blocks

Any prompt that has to work more than once should have all six. Missing blocks are where inconsistent output comes from.

01

Role

Who the model should act as.

02

Task

The job, in one clear line.

03

Context

The data, documents or facts it needs.

04

Constraints

Boundaries, rules, length limits.

05

Examples

One to five samples of the pattern you want.

06

Format

The exact shape of the output.

quarterly-summary
Role:      Senior financial analyst.
Task:      Summarise the Q3 report for the board.
Context:   <report> ...the actual figures... </report>
Limits:    Only figures from the report. Max 120 words.
Format:    3 bullets + a one-line risk callout.
ticket-triage
Role:      Support lead for an ERP rollout.
Task:      Classify this client ticket and route it.
Context:   Teams: finance, inventory, integrations.
Example:   "Invoice total wrong" → finance · high.
Format:    team · priority · one-line reason.
17
Prompt engineering04 · What goes wrong

Six mistakes that account for most bad output

Worth reading aloud to the team once. Every one of them is common, and every one is cheap to stop doing.

01

Everything in one mega-prompt

Build complexity through iteration. It is a conversation partner, not a search box.

02

Being vague

"Make it better" gives it nothing to aim at. Name the goal and the audience.

03

Stacking ten "do nots"

Prefer positive instructions. A pile of negatives degrades quality; use them sparingly.

04

Dumping everything in

More context is not better context. Curate high-signal material instead of drowning the task.

05

Magic filler phrases

"Take a deep breath" and similar folklore do not help. Clear structure does.

06

Iterating on vibes

Without a test set you cannot tell whether a change helped or just moved the error somewhere you did not look.

Vagueness: "write me a plan" returns a generic plan, because you never said the topic, the timeline or the goal.
Dumping: fifty pages pasted in and one question asked — the answer drowns, when the one relevant paragraph would have been enough.
18
Prompt engineering04 · From draft to reliable

How to refine a prompt

Start simple, then diagnose what is actually missing rather than guessing at rewordings.

01

Start plain

A clear, direct prompt with no examples yet. See what you get.

02

Diagnose the gap

Missing instruction? Missing context? Missing example? Name which one.

03

Add the matching fix

Examples for pattern. Steps for reasoning. A format spec for reuse.

04

Test and version

Run it against a small fixed set of cases. Keep what wins; record what you changed.

Translation: too formal → add "casual tone" → now too long → add "one short sentence". Three cycles, then it is reliable.
Email reply: too cold → paste one past reply you liked as an example → it now matches your voice without further instruction.

The habit that separates a team that improves from one that guesses

Even a twenty-example evaluation set is worth more than a week of impression-based tweaking. Loop until the prompt is reliable, not until it is impressive once.

19
05

How to actually
use it

The teams getting real output are not using better models. They are engineering the context the model works inside.

Practice05 · Three levels

Automation → agentic workflow → agent

Three different things, routinely confused, with very different costs and failure modes. Naming the level correctly is half of scoping a project.

Level 1 · rigid

Automation

Predefined, rule-based steps. The same input produces the same output every time. Ideal for high-volume, predictable work where correctness is non-negotiable.

When a form is submitted, send this email. Post the ledger entry on the first of the month.

Level 2 · hybrid

Agentic workflow

Deterministic rails orchestrate the flow, with model-powered steps dropped in at the genuinely fuzzy decision points. Predictable and flexible — the sweet spot for most business systems today.

Extract vendor data with a model, then route approval through fixed rules. Classify complaints, route each by existing policy.

Level 3 · autonomous

AI agent

The model decides actions at runtime. It reasons, picks tools, adapts, and persists until the goal is met rather than until a script ends. Most capable, least predictable, hardest to audit.

"Keep this client from churning" — it monitors, follows up, escalates.

Notice that almost everything valuable we deliver to clients sits at level 2. Level 1 is what we already build. Level 3 is where the demos live — and where the audit trail disappears.
21
Practice05 · The decision rule

Which one should you build?

One question decides it: can the logic be written down?

Build a workflow

When inputs are structured and the logic can be codified

And whenever you need an audit trail, regulatory defensibility, or a result that must be identical every single time. Payroll is the canonical case: the rules are fixed and the answer must be right on every run, not most runs.

Build an agent

When inputs are ambiguous and the logic resists codification

When the task genuinely needs interpretation and adaptation at runtime. Answering varied, unpredictable support tickets is the canonical case: every one is different and needs to be understood before it can be routed.

The rule of thumb

Start with the simplest option that works. Add agent reasoning only where branching logic would become unmaintainable. Most real systems end up combining both — the agent interprets the ambiguity, the workflow enforces the consistency — and that combination is usually the right answer rather than a compromise.

22
Practice05 · Context engineering

Prompt engineering context engineering

The unit of work stopped being the prompt. It is now the whole environment the agent sees: rules, tools, memory, scope, and the checks that run without being asked.

BASE RULES

AGENTS.md

What the project is, how it is built, what is forbidden

GUARDRAILS

Hooks

Lint, tests and blocks that run deterministically

REFERENCE

Skills

Reusable procedures loaded only when relevant

DELEGATION

Subagents

Isolated scopes, parallel work, clean contexts

TASK SCOPE

Spec

The acceptance criteria the work is checked against

Why this became necessary in 2026

Sessions now run for hours, so context has to survive compaction, overflow and restarts. And running five to fifty agents in parallel multiplies whatever your context setup is — good or bad.

The one-line version

Good context engineering is what makes agents consistently useful rather than occasionally impressive. Every hour spent here is repaid across every session afterwards.

23
06

Automation
platforms

Zapier, Make and n8n. The tools that actually build the level-one and level-two systems from the previous section — and the ones our clients will name.

Automation platforms06 · The category

What these platforms are, and why they matter to us

All three do the same fundamental job: connect software that was never designed to talk to each other, without writing and maintaining integration code. All three have now added AI and agent features on top.

The core job

Trigger → steps → action

Something happens in one system, a sequence of steps runs, something changes in another. A form submission creates an invoice. A payment updates the CRM. A file lands and gets filed.

The new layer

Model-powered steps

Each platform now lets you drop a reasoning step into the middle of that flow — classify this message, extract these fields, decide which team handles it — which is exactly the level-two pattern from section 05.

Why it is our problem

They compete with bespoke builds

A meaningful share of what we currently quote as custom integration work is four modules on one of these platforms. Knowing when that is true protects both our margin and our credibility.

The honest framing for a client conversation

These are not toys, and they are not a replacement for engineering either. They are the right tool when the work is plumbing between systems that already have APIs. They become the wrong tool when the logic gets deep enough that you are effectively programming in a visual editor with no tests and no version control.

25
Automation platforms06 · Zapier

Zapier — the broadest, the simplest, the priciest

The oldest of the three and still the default name. Its whole design philosophy is that a non-technical person finishes a working automation in minutes.

  • Around 7,000 integrations — more than any competitor, by a wide margin. If a client uses a niche regional CRM or an obscure industry tool, Zapier is the one most likely to already support it.
  • A linear builder, not a canvas. Pick a trigger app and event, pick an action, map the fields, activate. There is genuinely nothing to learn.
  • The AI layer: create workflows from a plain-language description, drop AI steps into a flow, and use Zapier Agents — assistants that work from instructions, browse, consult attached knowledge and act across connected apps, rather than following a fixed recipe.
  • Also ships a built-in database (Tables), simple front-ends (Interfaces), chatbots, and MCP access for connecting AI tools.
BillingHow it works
UnitThe task — roughly each successful step in a workflow. A ten-step workflow bills ten tasks per run.
Free~100 tasks a month, two-step workflows only — reduced sharply from earlier years.
Paid entryRoughly $20–30 a month for about 750 tasks, rising steeply from there.
AgentsA separate meter in "activities", with its own free allowance and paid tier. Agent usage does not draw on your task pool.
The two structural limits: it is the most expensive per unit of work at any real volume, and there is no self-hosted option — so it cannot answer a data-residency requirement.
26
Automation platforms06 · Make

Make — the visual middle ground

Formerly Integromat. A drag-and-drop canvas rather than a list of steps, which is what makes genuinely branching logic practical without writing code.

  • Several thousand integrations — fewer than Zapier, far more than n8n — plus routers, filters, iterators and error handling as first-class parts of the canvas.
  • Real logic without leaving the tool. Unlimited branching, and JavaScript or Python through a code module when the visual approach runs out.
  • AI agents since 2025, an AI toolkit, and connections to hundreds of AI services. You can use Make's own model provider, or bring your own key on the higher tiers — which matters if a client insists on a specific model or their own contract.
  • The sweet spot is operations and RevOps work: orchestrating dozens of SaaS tools with intelligent branching, without hiring a developer to maintain it.
BillingHow it works
UnitThe credit — one per module execution. A ten-step scenario costs ten credits per run. Make moved from "operations" to credits in 2025.
Free~1,000 credits a month, two active scenarios.
Paid entryAround $9–12 a month for roughly 10,000 credits — broadly three to five times cheaper than Zapier at equivalent volume.
The catchAI modules do not consume credits at the same rate as ordinary ones. This is where credit forecasts go wrong.
Known weaknesses: a real learning curve compared with Zapier, and debugging large scenarios with many branches and iterators remains the platform's historic soft spot.
27
Automation platforms06 · n8n

n8n — the one you can host yourself

A node-based canvas like Make, but source-available and self-hostable. For our market, that single property changes which conversations we can have.

  • Self-host the full product for free. Unlimited workflows, unlimited executions, every integration, on a server costing a few dollars a month. There is no usage meter at all.
  • Roughly 400–500 native integrations — far fewer than the others — but a generic HTTP node reaches any API, and JavaScript or Python runs inline inside nodes.
  • The strongest native AI stack of the three: dozens of nodes for agents, memory, vector stores and model calls, so retrieval and agent patterns are built in rather than bolted on.
  • European origin and in-perimeter hosting — workflow data never leaves infrastructure you control, which for a regulated client is sometimes the only acceptable architecture.
  • Billing differs in kind: the cloud tiers charge per workflow execution, not per step. A ten-step workflow run a thousand times is a thousand executions, where Zapier would bill ten thousand tasks.

The licence clause that catches agencies — including us

n8n is fair-code under the Sustainable Use License, not OSI open source. Self-hosting is free for your own internal business use. The moment you charge clients to run their automations on your instance, you need a paid commercial licence. The technical setup is identical either way, which is exactly why firms get this wrong.

Also worth knowing: the cost of "free" is real. Self-hosting means you own the server, the database, the backups, the upgrades and the security. That is a service we can sell — but it is not nothing, and it should be priced.
28