What people actually
complain about in AI agents

Every directory ranks AI agents by what they claim. This one starts from what users say goes wrong. Five recurring complaints, drawn from developer forums, vendor postmortems and analyst data — with the numbers attached and every source linked.

By category

Coding agentsReview burden, technical debt and the security rate nobody quotes.5 complaints, sourced →Support agentsEscalation loops, context loss, and the gap between deflection and resolution.5 complaints, sourced →Voice agentsLatency against a 300ms human expectation, and what interruption exposes.5 complaints, sourced →

Across every category

The agent gets worse and nobody tells you

Materially different behaviour month over month

The most repeated complaint is not that agents fail — it's that the same agent behaves differently month to month. Tasks that were routine start taking several attempts. Long refactors terminate early. Context windows quietly degrade and rate limits shift without notice.

What makes this rank first is the second half: vendors rarely communicate the change. Users discover a regression by losing an afternoon to it, which converts a technical problem into a trust problem.

If you're building one: Publish a changelog for model and prompt changes, not just features. Silence during a regression costs more goodwill than the regression.

dev.to — AI agents on Reddit, ten threads on cost and reliability ↗

Nobody can predict the bill

5–25× the cost of a standard chat interaction

One instruction can fan out into dozens of API calls, chained tools and silent retries. Buyers understand $49/month. They do not understand $0.004 per token plus $0.02 per retrieval call, and they cannot forecast either.

A classification task costing $0.01 in a chat interface runs $0.10–$0.50 once an agent handles it. An agent can clear a customer's entire credit balance in twenty minutes, and monthly reconciliation is useless against something running continuously.

If you're building one: Spend caps and a live cost meter do more for conversion than a cheaper rate. The fear being priced is the surprise, not the amount.

The AI Journal — the agentic AI billing problem ↗

Trust in full autonomy is collapsing

Trust in fully autonomous agents fell 43% → 27% in one year

This is the sharpest reversal in the data. Enterprises watched agents fail in real deployments and pulled back from autonomy rather than pushing further into it — the opposite of the direction most vendors are building.

Gartner expects 40% of enterprises to demote or decommission autonomous agents by 2027, over governance gaps found only after a production incident. The reliable deployments treat human oversight as a feature rather than an admission of failure.

If you're building one: Selling 'fully autonomous' now reads as a risk signal to the buyers with budget. Approval steps are a feature you can charge for.

Camunda — AI agents have a trust problem ↗

It cannot hold your codebase in its head

Fails on multi-file dependencies at ~50K lines

Agents handle isolated files well and fall apart on systems. Multi-file dependency chains in a large codebase are where the complaints concentrate — the agent makes a locally correct change that breaks something three files away.

The framing that keeps recurring is domain knowledge, not context length: an agent that doesn't know why the code is shaped this way makes decisions that are individually defensible and collectively wrong.

If you're building one: Retrieval quality beats window size. Being explicit about the size of codebase you handle well earns more trust than implying no limit.

dev.to — AI agents on Reddit, ten threads on cost and reliability ↗

No owner, no rollback, no audit trail

Visibility disappears at the human↔agent handoff

The complaint that keeps agents out of production isn't capability. It's that when one goes wrong at 3am there is no escalation path, no record of what it did, and no way to undo it.

Demos avoid what production adds: persistent state, unpredictable inputs, concurrency, permissions and recovery. Most agents that never ship were never unable to do the task — they were unaccountable while doing it.

If you're building one: An audit log and a rollback path are procurement requirements disguised as engineering chores. They unblock deals capability never will.

Gartner — uniform governance and AI agent failure ↗

Method

Findings are synthesised from published analyses of developer forums, vendor engineering posts and analyst research — not from scraped threads, so nothing here is presented as a verbatim user quote.
Every figure links to the source that reported it. Check them.
Complaints are ordered by how often they recurred across sources, not by severity.
Compiled August 2026. This is a snapshot and it will age.

Why this page exists

agentwar.lol ranks agents by what people pay. That's a deliberately shallow signal, so this page carries the other half: what users actually report.
The method here — finding the complaints people repeat and ranking them by intensity — is what UserConcern does across any market, not just this one.

What people actuallycomplain about in AI agents