A support agent that drafts a reply is table stakes now. The version worth paying for reads the ticket, pulls the order history, issues the refund, updates the CRM, and leaves a note explaining what it did and why. Same underlying model. Completely different software. One answers a question, the other closes a loop.
That gap is where the whole agent conversation is heading. The interesting question is no longer whether a model can handle one step of work. It’s whether a system can be trusted with twenty steps, a payment method, and a weekend of unsupervised runtime.
What “agentic” actually adds to the word AI
People use agentic to mean everything from a chatbot with one tool call to a system that plans its own week. The useful distinction is narrower. An agent has a goal, a set of tools, some record of what it has already tried, and the ability to decide the next move without being prompted at every turn. Remove the loop and the state, and you’re left with a very good autocomplete.
If you want the cleaner definition, this breakdown of what “agentic” actually adds to the word AI is worth ten minutes. The short version: the label describes delegated control, not model size. A small model with good tools and a tight loop will outperform a frontier model that has to ask permission for every keystroke.
Task horizons keep stretching, and that breaks old assumptions
In March 2025, researchers at METR published an analysis showing that the length of task an AI agent could complete at roughly 50% reliability had been doubling about every seven months. At the time, that meant tasks worth maybe fifty minutes of human effort. Extend the trend two more years and you’re talking about work measured in days.
Curves aren’t destiny, and this one has obvious limits. Long tasks compound errors. A 95% success rate per step becomes close to a coin flip across twenty steps. Still, the direction matters enormously. Agents built for a single exchange fall apart when a job runs overnight. Agents built for persistence start to look less like features and more like colleagues who need onboarding.
Memory is the bottleneck nobody wants to talk about
Ask anyone who has pushed an agent into production what broke first. You rarely hear about model quality. You hear about state. The agent forgot a constraint from three days ago, re-read the same file six times, or contradicted a decision it made on Tuesday.
Bigger context windows help, but a bigger window is a bigger desk, not a memory. It fills up, it gets expensive, and relevance decays with every token you cram in. There’s a good practical rundown of why AI agents forget and what actually fixes it, and the fixes tend to be unglamorous plumbing: summarising old turns, writing decisions to a structured store, retrieving only what the current step needs, and deliberately dropping anything stale.
Teams that solve this quietly get a compounding advantage. Their agents finish tasks. Everyone else’s agents restart them.
Agents are starting to work in teams
A single agent hits a ceiling fast: one context, one perspective, one pass at reviewing its own work. Multi-agent setups split the job across roles with different instructions and different context windows. A planner hands a spec to a builder, the builder hands code to a reviewer, and the reviewer’s objections go back upstream.
MetaGPT is a decent illustration of the pattern. You describe a small tool, and a simulated team of product manager, architect, and engineer agents produces requirements, designs, and code in sequence. You can build a small Python tool with a team of AI agents in an afternoon and watch the hand-offs break in genuinely instructive ways.
The honest verdict so far: multi-agent systems shine when roles are truly separable and a clean artifact passes between them. They flounder when every agent shares the same context and simply argues. Coordination has a cost, and token bills multiply faster than quality does.
Real systems are already doing the work
This stopped being a research curiosity some time ago. Look at deployments rather than launch videos. The pattern across ten real AI agent systems doing actual work is strikingly consistent: narrow scope, strong tools, and a human sitting at the approval gate.
- Coding agents that triage an issue, write the patch, run the test suite, and open a pull request for review.
- Finance agents reconciling invoices against purchase orders and flagging the 3% that don’t match.
- Security agents triaging alerts and closing obvious false positives before an analyst ever sees them.
- Sales ops agents enriching records and drafting follow-ups from call transcripts.
Notice what these share. Measurable output, a clear stopping point, and a human who can reject the work cheaply. Agents that fail one of those three tests tend to end up back in demo land.
Permissions, blast radius, and a bill that arrives later
An agent with write access is an attack surface with a personality. Anthropic has publicly warned that small AI models could self-replicate through IoT devices, which reads like science fiction until you remember that an agent’s tool list is precisely the capability set an attacker wants. Email sending, file deletion, API keys, shell access. That’s the prize.
The guardrails are boring and effective. Least privilege per tool, per-action confirmation on anything irreversible, rate limits on outbound actions, and an audit log a human can actually read at 2am. Most incidents won’t come from a rogue intelligence. They’ll come from a well-meaning agent with slightly too much scope and no circuit breaker.
Cost per completed task is the number that matters
Token prices keep falling, but agent consumption rises faster. One task can burn 50,000 to 500,000 tokens once you count retries, tool outputs, and reflection loops. That makes cost per million tokens close to meaningless as a metric. Cost per completed task is the one a finance team will ask about.
It reshapes design too. Cheap models for triage and routing, expensive models reserved for the hard step, caching for repeated lookups, hard caps on loop iterations. An agent that costs four dollars to finish work a person does for twelve is a business. One that costs forty and still needs a retry is a hobby with an API key.
The unglamorous work that will decide who wins
Almost none of the next two years will be won on model selection. It’ll be won on evals, observability, and rollback. Can you replay last Tuesday’s failed run and see exactly which step went wrong? Can you diff this week’s agent behaviour against last week’s? Can you kill a runaway job in one click without losing the partial output?
Teams answering yes to those questions are shipping agents that hold up under real workloads. Teams that can’t answer them are one bad tool call away from a very long week. Start with a task that’s measurable, give the agent the smallest toolset that gets the job done, log every action, and put a human at the one gate where a mistake would be expensive. Everything else is a detail you can tune later.