AI Agents

Gartner's 2026 CIO and Technology Executive Survey, cited in its Hype Cycle for Agentic AI (15 April 2026), found 17% of organisations have deployed AI agents and more than 60% expect to within two years (as reported by Digital Applied, July 2026).

Deloitte's Tech Trends 2026 (10 December 2025) reported 11% of organisations with production-ready agentic systems and 42% without a formal agentic AI strategy.

McKinsey's State of AI (November 2025) found 62% of organisations at least experimenting with agents and 23% scaling an agentic system somewhere, with no more than 10% scaling agents in any single business function.

Gartner's June 2025 press release predicts over 40% of agentic AI projects cancelled by end-2027 on cost, unclear value or weak risk controls. Gartner also projects 33% of enterprise software apps will include agentic AI by 2028, up from under 1% in 2024.

Grand View Research puts the AI agents market at $7.63B in 2025 and $10.91B in 2026; other firms (Global Market Insights, MarkNtel) give about $10-11B for 2026, with wide long-range variation.

Not refreshed: Google AI-generated code share, developer daily-use rate, the MIT coding study figures, and the Gartner 2028 and 2035 forecasts could not be re-verified and are left as written.

1. Goal: you give it an outcome, not instructions.
2. Plan: it breaks the goal into steps.
3. Act: it uses tools (browsers, code, APIs, files).
4. Check & adapt: it verifies results, fixes errors, and continues.

Even if only half-right, delegated decision-making is entering the workplace faster than any prior enterprise technology.

1. 2023AutoGPT: the viral, broken first draft
2. 2024Computer use: models learn to click and type
3. 2025Protocols: MCP standardises agent-tool wiring
4. 2026Embedded agents: 40% of enterprise apps (Gartner)
5. 2028Delegation era: 15% of decisions autonomous

1. Guardrails: permissions, approvals and audit trails.
2. Orchestration: the harness that runs the loop and enforces limits.
3. Memory: context it keeps across steps and sessions.
4. Tools: what it can touch, including code, browsers, databases, apps.
5. The model: the reasoning engine that plans and decides.

• Reasoning models: models that think before acting, cutting error rates.
• Tool protocols: standards like MCP let models drive real software.
• Long context: agents can hold a whole task in working memory.
• Computer use: models that see screens and click like a person.
• Cheap inference: thousands of steps per task became affordable.
• Verification loops: agents that test their own work before shipping it.

An agent's real cost isn't the model, it's the error rate compounding across every step of the task. A 99%-reliable step fails 1 time in 100. Chain 50 steps and the task fails 4 times in 10.

Software engineering is the first industry visibly reorganised by agents, because code can be tested, checked and rolled back.

Where errors are cheap to catch, autonomy scales fast. That is why coding went first.

• Customer operations: resolving tickets end-to-end, not just drafting replies.
• Research & analysis: multi-source investigation compressed from days to minutes.
• Back office: invoices, claims, reconciliation; high-volume, rule-rich work.
• Sales ops: CRM hygiene, outreach drafts, pipeline management.
• IT & security: triaging alerts, patching, access reviews.
• Personal work: inboxes, scheduling, bookings; slower, higher-trust.

• Anthropic: Claude and Claude Code, the reasoning-and-coding agent leader.
• OpenAI: consumer agents at ChatGPT scale; operator-style computer use.
• Microsoft: Copilot embedded across the enterprise estate.
• Salesforce / SaaS: every incumbent bolting agents onto its workflow.
• Startups: thousands of vertical agents for support, legal, finance, ops.
• Open source: frameworks and local agents keeping pricing honest.

A two-by-two map of value per task (vertical) against ease of deployment (horizontal). Coding and research sit top-right, the frontier, with customer ops just right of centre. Legal work and healthcare admin sit top-left (hard, rich). Back office sits near the centre line, on the easy side. The bottom quadrants (hard/modest and easy/modest) are empty.

If even directionally right, agents are not a product category; they are the next business model of the software industry.

Gartner expects over 40% of agentic AI projects to be cancelled by 2027, not because the models fail, but because the organisations do.

Unclear ROI, missing governance, and workflows that were never redesigned for delegation.

• Demo-to-production gap: 95% reliability wows a boardroom and breaks a workflow.
• No owner: agents cross team boundaries; nobody owns the outcome.
• Unbounded cost: loops that burn tokens without a budget or kill-switch.
• Compliance stall: legal can't sign off on what it can't audit.
• Workflow mismatch: automating a broken process just breaks it faster.
• Vendor-washing: 'agents' that are chatbots with a new label.

• Narrow first: one well-bounded workflow, measured end to end, not a general assistant.
• Verification built in: every action checkable, reversible, or approved.
• A human escalation path: agents handle the routine; people own the exceptions.
• Real baselines: measured against the process's true cost, not a vendor's slide.

1. Read-only: agents observe and draft; humans act.
2. Approve-to-act: agents act only on explicit sign-off.
3. Act-and-log: routine actions execute, fully audited.
4. Autonomous: trusted workflows run unsupervised, with limits.

• Prompt injection: malicious content hijacking an agent's instructions.
• Cascading errors: one wrong step compounding through a chain.
• Data exposure: agents with broad access leaking what they touch.
• Runaway cost: loops without budgets or limits.
• Audit gaps: actions no one can reconstruct after the fact.
• Over-delegation: handing off judgement that was the business's edge.

The winning formula of the agent era is not full autonomy, it is human judgement, multiplied by tireless machine execution.

The centaur, not the robot: people setting direction, agents doing the legwork, systems checking the work.

• Production %: real deployments, not pilots or claims.
• Task success: reliability, the 95 to 99.9 climb.
• $/task: unit economics, falling with compute.
• Incidents: failures and breaches, the trust ledger.

1. Does production catch up to pilots? The 11% number is the one to watch.
2. Does reliability cross the threshold? Unsupervised operation in real workflows.
3. Does outcome pricing stick? Per-task pricing rewires software economics.
4. Does a public failure set rules? One big incident could define regulation.
5. Do agents talk to agents? Inter-agent commerce would be a new economy.

Agents will feel disappointing for two more years and then suddenly indispensable, the standard arc of every platform shift.

The web in 1997, mobile in 2009, cloud in 2012: mocked as toys, then abruptly the default. Agents are on the same clock.

The chatbot era asked what AI knows. The agent era asks what AI may do, and every hard question in business, from trust to accountability, hides inside that single word: may.

1. Action is the product: agents sell finished work, not answers.
2. Reliability is the moat: the 95 to 99.9 climb is the whole game.
3. Coding is the preview: every industry gets its version, in verifiability order.
4. Most projects will fail: organisation, not capability, is the constraint.
5. Own the workflow: value accrues to whoever owns the work being automated.

• Gartner: agentic AI forecasts, Hype Cycle and CIO survey (2025-26).
• MIT: study of AI agents across 100,000+ developers (2026).
• Google / Microsoft: disclosures on AI-generated code share.
• Stack Overflow / developer surveys: AI tool adoption and trust (2025-26).
• Industry forecasts: AI agents market sizing (2025-26).
• Anthropic / OpenAI: agent product documentation and safety frameworks.
• Figures marked 'estimate' or 'projection' are compilations; this field moves fast.