TRENDING

TECHNOLOGY NEWS FROM AROUND THE WORLD

→
Looking back at 2025: AI agents and the difference between chatting and acting
Artificial Intelligence · 2025Photo: Unsplash / Unsplash ↗
Artificial Intelligence2025 / RETROSPECTIVE

Looking back at 2025: AI agents and the difference between chatting and acting

As tool-using agents attracted attention, the central engineering challenge became defining their permitted actions and operational limits.

Yazılımcının GünlüğüPublished: Nov 15, 20263 min read

Sample period date: Jun 16, 2025. These essays revisit earlier technologies from today's editorial perspective; they are not represented as posts originally published in those years.

From conversation to action

A model that answers questions and an agent that edits files or updates production records do not pose the same risks. Interest grew around tool access, planning and task execution. These capabilities brought new failure cases as well as productivity opportunities.

Trust must be measurable

Counting agent steps does not show progress. Measure verified completions, error rates, retry cost and rollback ability. When multiple agents update the same data, uniqueness constraints and authorization need explicit protection.

Conceptual workflow — original editorial illustration

A principle that still applies

Agents are most valuable when they reduce repetitive work without removing accountability. Deletion, payment and external publishing should remain behind authorization checks. Intelligent automation still needs understandable logs and verifiable outcomes.

From conversation to action around 2025

Developer attention shifted from systems that answered questions toward agents that could edit files, operate tools and update records. This shift did not automatically imply safer or more intelligent software. An agent needed to choose the correct resource, test its modifications and recover safely from failure. The impact of an error increases with the permission to act. Evaluation should therefore focus less on conversational style and more on permitted operations and verifiable outcomes.

Avoid duplicate records during retries

An agent may retry a request after a timeout, but the original operation might already have succeeded. A second attempt could create another invoice, booking or comment. Stable operation identifiers, unique database constraints and idempotent endpoints are crucial. Retrying based solely on a missing HTTP response can damage data. The hardest reliability problems appear in these uncertain states, where success and failure cannot be inferred from one network event.

Measure operational success

Track verified task completions, human interventions, latency, cost and recoverable failures. Many tool calls do not constitute productive work. Set concurrency based on measured API behavior and error rates. Protect irreversible deletion, payments and publishing behind separate authorization policies. Explainable action logs and verification receipts are essential if users are expected to trust autonomous progress. This essay evaluates the ideas of 2025 from a later editorial perspective.

Further reading / Referencehttps://github.blog/ai-and-ml/ ↗
0 reads

Comments

0 real comments
FICTIONAL EXAMPLELeyla · 2025 (sample)

An agent retry should not record the same action twice. Idempotent tool operations and clear error logs make that possible.

Illustrative reader note and date; not a real comment.
✎ Write a comment

Comments are reviewed before publication.