Od klepetalnih robotov do sodelavcev: zakaj se naslednji val umetne inteligence osredotoča na upravljanje agentov, ne na pogovor z njimi

/Splošno/ Avtor

Od klepetalnih robotov do sodelavcev: zakaj se naslednji val umetne inteligence osredotoča na upravljanje agentov, ne na pogovor z njimi

V zadnjih nekaj letih je večina ljudi umetno inteligenco doživljala kot enega samega sogovornika: vnesete poziv, model odgovori in vi ponovite. Ta vzorec interakcije je znan in pogosto uporaben – vendar doseže tudi zgornjo mejo. Pravo delo v podjetjih le redko izgleda kot ena sama nit vprašanj in odgovorov. Izgleda kot neurejena mreža dokumentov, kodnih baz, čakalnih vrst zahtevkov, skupnih map, odobritev, dovoljenj in napol dokončanih nalog, ki jih je treba uskladiti.

Zato se dogaja opazen premik: prodajalci umetne inteligence vas poskušajo premakniti izklepet z modelomdonadzorovanje nabora modelov— majhna »ekipa« agentov, ki si lahko razdelijo delo, delujejo vzporedno, ohranjajo svoj kontekst in poročajo.

Objave podjetij Anthropic in OpenAI tega tedna so jasen znak tega preobrata. Anthropic skupaj z novim vodilnim modelom, Claude Opus 4.6, uvaja »ekipe agentov« znotraj Claude Code. OpenAI pa promovira Frontier, platformo, namenjeno gradnji, uvajanju in upravljanju »sodelavcev umetne inteligence« znotraj podjetij.

Predstavitev je zapeljiva: postanete vodja, agenti postanejo izvajalci. Toda realnost je bolj zapletena. V praksi delo ne izgine – spremeni se. Človek se iz delavca spremeni v tistega, kivodja, in upravljanje ima svoje stroške: nadzor, koordinacijo, nadzor kakovosti in odgovornost, ko gre kaj narobe.

Kaj ljudje mislijo z "agenti" (in zakaj to ni samo trženje)

V vsakdanjem pogovoru se »agent« pogosto uporablja kot nejasen sinonim za »umetno inteligenco, ki lahko počne stvari«. Tehnično gledano večina sodobnih agentov združuje nekaj sestavin:

  • Zanka načrtovanja: sistem se namesto enkratnega odgovora večkrat odloči, kaj storiti.
  • Uporaba orodja: agent lahko kliče API-je, izvaja ukaze, bere/piše datoteke, brska po notranjih bazah znanja ali odpira zahteve.
  • Stanje in spomin: agent vodi lokalno evidenco o tem, kaj je storil, kaj se je naučil in kaj je še treba narediti.
  • Dovoljenja: neka plast odloča, do česa ima agent dostop in katera dejanja lahko izvede.

Če ste že uporabljali pomočnika za kodiranje, ki lahko izvaja teste, odpira zahteve za prevzem ali išče po repozitoriju, ste že videli razliko med klepetalnim robotom in agentom, ki uporablja orodja. Ključno je, da agent ni leustvarjanje besedila— jedelovanje znotraj delovnega toka.

Ideja »tima« doda še eno plast: namesto ene zanke imate več zank, ki se izvajajo hkrati. En agent lahko pregleda kodno bazo. Drug lahko piše dokumentacijo. Tretji lahko išče robne primere ali se loti problema z drugačno hipotezo. V idealnem primeru se zbližajo hitreje kot en sam asistent.

Anthropic: ekipe agentov v Claude Code

Anthropicov novi model,Claude Opus 4.6, je pozicioniran kot močnejši model »opravljanja težkih stvari«: izboljšano kodiranje, boljše odpravljanje napak in pregledovanje, daljše vzdrževanje nalog in beta kontekstno okno, ki lahko doseže1 milijon žetonovTa zadnja točka je pomembna, ker je veliko napak "agentov" v resnici napak "konteksta" – agenti z naraščanjem števila nalog izgubijo sled za podrobnostmi.

Poleg modela se Anthropic nagiba k razvijalski uporabniški izkušnji:ekipe agentovznotraj Claude Code. Osnovna ideja je preprosta:

  • Začneš ekipo.
  • Ena seja deluje kot vodilo.
  • Več sej soigralcev deluje neodvisno v svojih kontekstih.
  • Soigralci si lahko pošiljajo sporočila neposredno (ne samo prek vodje).

V lastni dokumentaciji Anthropic poudarjajo, da ekipe najbolje delujejo, kadar je delo mogoče čisto vzporedno izvajati: raziskave, pregledi, modularne spremembe, konkurenčne hipoteze za odpravljanje napak, delo med plastmi, kjer lahko različni agenti posedujejo različne površine.

Vendar dokumentacija priznava tudi skrito resnico: ekipe stanejo več. Porabijo več žetonov in dodajo stroške koordinacije. Če je vaša naloga zaporedna ali globoko prepletena (na primer refaktoriranje, ki se dotika številnih skupnih datotek), lahko ena dobra seja premaga »ekipo«, ki nenehno stopa čez samo sebe.

OpenAI-jev pogled: Frontier kot ravnina nadzora agentov

OpenAIMejaje manj namenjen posameznim razvijalcem in bolj organizacijam. Bere se kot »upravljalna plast« za uvajanje agentov:

  • Skupni kontekst med poslovnimi sistemi (podatkovna skladišča, CRM-ji, interne aplikacije).
  • Identiteta agenta, dovoljenja in meje.
  • Dosledno izvedbeno okolje.
  • Povratne zanke in vrednotenje za izboljšanje agentov.

To je pogosta težava podjetij, odeta v jezik umetne inteligence: podjetja se že tako soočajo s širjenjem, upravljanjem in razdrobljenimi sistemi. Če vsaka ekipa uporablja agente, ki se povezujejo z različnimi orodji z različnimi dovoljenji, je rezultat kaos. Frontier je postavljen kot način za centralizacijo tega upravljanja.

Sporočilo je tudi konkurenčno. Ponudniki SaaS so zgodovinsko gledano prodajali programsko opremo, ki koordinira poslovne poteke dela. Če lahko vzorčna podjetja zapakirajo »poteke dela kot agente«, lahko dosežejo neposredno to področje.

Neprijetna realnost srednjega managementa

To je del, ki ga prodajalci redko poudarijo: človeška vloga se spremeni iz »delavca« v »nadzornika«. To se sliši opolnomočujoče, vendar prinaša kompromise.

1) Zdaj imate v lasti nadzor kakovosti

Agenti lahko hitro pišejo osnutke, vendar še vedno halucinirajo, napačno berejo kontekst ali izvajajo verjetne, a napačne spremembe. Ko zaženete enega agenta, pregledujete en tok izhodnih podatkov. Ko zaženete pet agentov, pregledujete pet.

Z drugimi besedami: ekipa agentov lahko pomnoži vašo produktivnost, lahko pa tudi pomnoži površino, kjer lahko gre kaj narobe.

2) Stroški koordinacije so resnični

Vzporednost deluje, ko so naloge neodvisne. Toda veliko resničnih projektov je povezanih:

  • Sprememba podatkovnega modela sili spremembe povsod.
  • Popravek napake se nanaša na kodo, teste, dokumentacijo in uvajanje.
  • Sprememba politike se dotika usposabljanja, komunikacije in orodij.

Tudi v človeških ekipah lahko vzporedno delo povzroči konflikte združevanja in neusklajenost. Pri agentih se te težave pojavijo hitreje, ker agenti delujejo hitro – in ne zaznajo naravno organizacijskega konteksta, razen če ga izrecno navedete.

3) Odgovornost ne izgine

Če agent odpre zahtevo za prevzem (pull request), ki prekine produkcijo, ne morete reči »umetna inteligenca je to storila«. Oseba, ki jo je odobrila, je lastnica rezultata. Podjetja se zanimajo za dovoljenja in revizijske sledi, ker je dejanje agenta še vedno korporativno dejanje.

Zato izraz »sodelavci agenta« hitro pripelje do vprašanj o:

  • Kdo je agentu odobril dostop?
  • Katere podatke je prebral?
  • Katere ukrepe je sprejelo?
  • Ali lahko reproduciramo njegovo verigo sklepanja?
  • Ali lahko varno razveljavimo spremembe?

Zakaj se premik dogaja zdaj

Ta val ni naključen. Nekaj ​​trendov se je združilo.

Daljši konteksti naredijo agente manj krhke

Če lahko model resnično zanesljivo sledi sto tisočem žetonov, lahko v delovnem pomnilniku shrani več kodne baze, zgodovine zahtevkov ali priročnika s pravilniki. To zmanjša napake »izgubljene niti«, zaradi katerih so bili prejšnji agenti frustrirajoči.

Ekosistemi orodij dozorevajo

V letih 2023–2024 so se številni predstavitveni primeri agentov zanašali na krhko avtomatizacijo brskalnikov ali na nabore orodij za igrače. Danes so podjetja zgradila resnične integracije: dostop do repozitorij, delovna drevesa, sisteme za izdajanje zahtevkov, notranje iskanje, dostop do strukturiranih podatkov in izvajalna okolja v peskovniku.

Razlika med agentom, ki lahkogovoriti o popravkuin takšen, ki lahkoizvedite teste in odprite PRje ogromno.

Podjetja si želijo vzvoda, ne novosti

Klepetalni roboti so uporabni, vendar so pogosto aditivni: ljudem pomagajo, da iste naloge opravijo nekoliko hitreje. Agentske platforme poskušajo biti multiplikativne: dejansko poskušajo človeku prevzeti naloge.

Tu živi proračun. Če lahko podjetje šesttedensko optimizacijsko nalogo spremeni v en dan (kot trdi OpenAI v Frontier marketingu), to ni »kul tehnologija« – to je strateška prednost.

Kako izgleda »dobro« za ekipe agentov

Če eksperimentirate z večagentnimi nastavitvami, najbolj zanesljive zmage običajno izhajajo iz nekaj vzorcev.

Delite po artefaktu, ne po nejasni vlogi

»Raziskovalec, programer, preizkuševalec« je preveč nejasno. Boljše delitve izgledajo takole:

  • Agent A: pregleda repozitorij in poišče, kje se uporablja X; ustvari seznam s potmi datotek.
  • Agent B: pripravi načrt migracije; naštej tveganja in predlagaj teste.
  • Agent C: spremembe izvedite samo v modulu Y; drugih modulov se ne dotikajte.
  • Agent D: pisanje opomb ob izdaji in posodabljanje dokumentacije.

Betonske meje zmanjšujejo prekrivanje in olajšajo pregled.

Kontrolne točke sile

Varen vzorec je:

  1. Agent predlaga načrt.
  2. Človek odobri ali popravi.
  3. Agent se izvaja v omejenem obsegu.
  4. Človeški pregledi in združitve.

To je počasneje kot "pustiti, da teče", vendar je veliko bolj zanesljivo.

Z agenti ravnajte kot s pripravniki z nadnaravnimi močmi

Dober miselni model je: agenti so hitri, neutrudni in imajo široko znanje – vendar jim manjka presoje na področjih, ki so najpomembnejša. Lahko so briljantni pri ustvarjanju možnosti in grozni pri izbiri tiste, ki ustreza vašim dejanskim omejitvam.

Če z njimi ravnate kot z avtonomnimi sodelavci, se boste opekli. Če z njimi ravnate kot z visoko produktivnimi pomočniki, ki potrebujejo vodenje in pregled, boste dobili resnično vrednost.

Tveganja, ki bodo zaznamovala naslednje leto

Z širjenjem platform agentov se bodo nekatera tveganja povečala.

Širjenje varnosti in dovoljenj

Več orodij kot lahko agent uporabi, bolj nevarno postane posamezno vbrizgavanje poziva ali napačna konfiguracija. Podjetja bodo vse bolj zahtevala dovoljenja z najmanjšimi privilegiji, izvajanje v peskovniku in močne dnevnike nadzora.

Vrednotenje in metrike »zanesljivosti agentov«

Primerjalni testi za zmogljivost surovega modela so ena stvar. Primerjalni testi za to, »ali lahko ta agent dokonča delovni tok v vašem okolju, ne da bi pri tem pokvaril stvari«, so nekaj drugega. Pričakujte val orodij za ocenjevanje agentov: ponovljive zagone, točkovanje, sledenje regresiji in avtomatizirano združevanje v rdeče skupine.

Zaklepanje delovnega toka

Če podjetje gradi svoje delovanje okoli izvajalnega okolja in kontekstne plasti agenta enega prodajalca, se stroški prehoda povečajo. To bo spodbudilo povpraševanje po standardih in prenosljivosti – vendar bodo prodajalci konkurirali tudi tako, da bodo svoje ekosisteme naredili vezane.

Bistvo

Umetna inteligenca se premika izodgovarjanjedoigralstvoin izen asistentdoekipe agentovPrednost je resnična: vzporedno delo, hitrejše raziskovanje in možnost delovanja znotraj neurejenih sistemov resničnega sveta namesto v klepetalnici.

Vendar ta premik ne odstrani ljudi iz zanke – spremeni jo. Če se zgodba o »sodelavcu z umetno inteligenco« uresniči, bodo mnogi znanstveni delavci porabili manj časa za pripravo prvih osnutkov in več časa za nadzor, revidiranje in usklajevanje. Naslednje bojišče produktivnosti ne bo to, kdo ima najpametnejšega klepetalnega robota. Bo to, kdo lahko agente naredi zanesljive, obvladljive in resnično uporabne znotraj resničnih organizacij.

Viri

Document Title
From chatbots to co-workers: why the next wave of AI is about managing agents, not talking to them
AI vendors are shifting from single chatbots to teams of tool-using agents. Here’s what agent teams are, why they’re emerging now, and why supervision and governance matter.
Title Attribute
oEmbed (JSON)
oEmbed (XML)
JSON
View all posts by Admin
Bing delisted 1.5 million Neocities sites. Here’s why that matters for the open web
The “Virtual Console” is back… sort of: what Nintendo’s Switch 2 retro downloads really mean
Page Content
From chatbots to co-workers: why the next wave of AI is about managing agents, not talking to them
Nature
Climate
/
General
/ By
Admin
For the last couple of years, most people have experienced AI as a single conversational partner: you type a prompt, the model responds, and you iterate. That interaction pattern is familiar and often useful — but it also hits a ceiling. Real work inside companies rarely looks like a single Q&A thread. It looks like a messy web of documents, codebases, ticket queues, shared folders, approvals, permissions, and half-finished tasks that have to be coordinated.
That’s why a noticeable shift is underway: AI vendors are trying to move you from
chatting with a model
to
supervising a set of models
— a small “team” of agents that can split work, run in parallel, keep their own context, and report back.
This week’s announcements from Anthropic and OpenAI are a clear sign of that pivot. Anthropic is pushing “agent teams” inside Claude Code alongside a new flagship model, Claude Opus 4.6. OpenAI is pushing Frontier, a platform aimed at building, deploying, and governing “AI coworkers” inside enterprises.
The pitch is seductive: you become the lead, agents become the doers. But the reality is more complicated. In practice, the job doesn’t disappear — it changes. The human moves from being the worker to being the
manager
, and management has its own costs: oversight, coordination, quality control, and responsibility when something goes wrong.
What people mean by “agents” (and why it’s not just marketing)
In everyday conversation, “agent” is often used as a vague synonym for “AI that can do things.” Technically, most modern agents combine a few ingredients:
A planning loop
: the system repeatedly decides what to do next instead of answering once.
Tool use
: the agent can call APIs, run commands, read/write files, browse internal knowledge bases, or open tickets.
State and memory
: the agent keeps a local record of what it has done, what it learned, and what still needs doing.
Permissions
: some layer decides what the agent is allowed to access and what actions it can take.
If you’ve used a coding assistant that can run tests, open pull requests, or search a repo, you’ve seen the difference between a chatbot and a tool-using agent. The key point is that the agent is not only
generating text
— it is
operating inside a workflow
.
The “team” idea adds another layer: instead of one loop, you have multiple loops running at once. One agent can scan a codebase. Another can draft documentation. A third can chase down edge cases or attack the problem from a different hypothesis. Ideally, they converge faster than a single assistant could.
Anthropic’s take: agent teams in Claude Code
Anthropic’s new model,
Claude Opus 4.6
, is positioned as a stronger “do hard things” model: improved coding, better debugging and review, longer sustained tasks, and a beta context window that can reach
1 million tokens
. That last point matters because a lot of “agent” failures are really “context” failures — agents lose track of details as tasks grow.
Alongside the model, Anthropic is leaning into a developer UX:
agent teams
inside Claude Code. The basic idea is straightforward:
You start a team.
One session acts like a lead.
Multiple teammate sessions work independently in their own contexts.
Teammates can message each other directly (not only through the lead).
Anthropic’s own docs emphasize that teams work best when work can be parallelized cleanly: research, reviews, modular changes, competing debugging hypotheses, cross-layer work where different agents can own different surfaces.
But the docs also admit the hidden truth: teams cost more. They burn more tokens and they add coordination overhead. If your task is sequential or deeply entangled (for example, a refactor that touches many shared files), one good session may beat a “team” that is constantly stepping on itself.
OpenAI’s take: Frontier as the agent control plane
OpenAI’s
Frontier
is aimed less at individual developers and more at organizations. It reads like a “management layer” for agent deployment:
Shared context across business systems (data warehouses, CRMs, internal apps).
Agent identity, permissions, and boundaries.
A consistent execution environment.
Feedback loops and evaluation so agents improve.
This is a common enterprise problem dressed in AI language: companies already struggle with sprawl, governance, and fragmented systems. If every team spins up agents that connect to different tools with different permissions, the result is chaos. Frontier is positioned as a way to centralize that governance.
The message is also a competitive one. SaaS vendors historically sell software that coordinates business workflows. If model companies can package “workflows as agents,” they can reach directly into that territory.
The uncomfortable middle-manager reality
Here’s the part vendors rarely emphasize: the human role changes from “worker” to “supervisor.” That sounds empowering, but it comes with tradeoffs.
1) You now own quality control
Agents can draft quickly, but they still hallucinate, misread context, or make plausible-but-wrong changes. When you run one agent, you’re reviewing one stream of output. When you run five, you’re reviewing five.
In other words: a team of agents can multiply your output, but it can also multiply the surface area where things can go wrong.
2) Coordination overhead is real
Parallelism works when tasks are independent. But many real projects are coupled:
Changing a data model forces changes everywhere.
A bug fix touches code, tests, documentation, and deployment.
A policy change touches training, communication, and tooling.
Even in human teams, parallel work can create merge conflicts and misalignment. With agents, those issues appear faster because agents work quickly — and they don’t naturally “sense” organizational context unless you explicitly provide it.
3) Accountability doesn’t go away
If an agent opens a pull request that breaks production, you don’t get to say “the AI did it.” The person who approved it owns the outcome. Enterprises care about permissions and audit trails because the agent’s action is still a corporate action.
This is why “agent coworkers” quickly leads to questions about:
Who granted the agent access?
What data did it read?
What actions did it take?
Can we reproduce its chain of reasoning?
Can we roll back the changes safely?
Why the shift is happening now
This wave is not random. A few trends have converged.
Longer contexts make agents less brittle
If a model can genuinely track hundreds of thousands of tokens reliably, it can hold more of a codebase, a ticket history, or a policy manual in working memory. That reduces the “lost the thread” failures that made earlier agents frustrating.
Tool ecosystems are maturing
In 2023–2024, many agent demos relied on brittle browser automation or toy toolsets. Today, companies have built real integrations: repo access, worktrees, ticketing, internal search, structured data access, and sandboxed runtimes.
The difference between an agent that can
talk about a fix
and one that can
run tests and open a PR
is enormous.
Businesses want leverage, not novelty
Chatbots are useful, but they’re often additive: they help people do the same tasks a bit faster. Agent platforms are trying to be multiplicative: they try to actually take tasks off the human’s plate.
That’s where the budget lives. If a company can turn a six-week optimization task into one day (as OpenAI claims in Frontier marketing), that’s not “cool tech” — it’s a strategic advantage.
What “good” looks like for agent teams
If you’re experimenting with multi-agent setups, the most reliable wins tend to come from a few patterns.
Divide by artifact, not by vague role
“Researcher, coder, tester” is too fuzzy. Better splits look like:
Agent A: scan the repo for where X is used; produce a list with file paths.
Agent B: draft a migration plan; list risks and propose tests.
Agent C: implement changes only in module Y; do not touch other modules.
Agent D: write release notes and update docs.
Concrete boundaries reduce overlap and make review easier.
Force checkpoints
A safe pattern is:
Agent proposes a plan.
Human approves or corrects.
Agent executes in a constrained scope.
Human reviews and merges.
This is slower than “let it run,” but it’s far more dependable.
Treat agents like interns with superpowers
A good mental model is: agents are fast, tireless, and broadly knowledgeable — but they lack judgment in the ways that matter most. They can be brilliant at generating options and terrible at choosing the one that fits your actual constraints.
If you treat them like autonomous coworkers, you’ll get burned. If you treat them like high-output assistants that need guidance and review, you’ll get real value.
The risks that will define the next year
As agent platforms spread, a few risks will become sharper.
Security and permissions sprawl
The more tools an agent can use, the more dangerous a single prompt injection or misconfiguration becomes. Enterprises will increasingly demand least-privilege permissions, sandboxed execution, and strong audit logs.
Evaluation and “agent reliability” metrics
Benchmarks for raw model capability are one thing. Benchmarks for “can this agent complete a workflow in your environment without breaking things” are another. Expect a wave of agent evaluation tooling: replayable runs, scoring, regression tracking, and automated red-teaming.
Workflow lock-in
If an enterprise builds its operations around one vendor’s agent runtime and context layer, switching costs rise. This will push demand for standards and portability — but vendors will also compete by making their ecosystems sticky.
Bottom line
AI is moving from
answering
acting
, and from
one assistant
teams of agents
. The upside is real: parallel work, faster exploration, and the ability to operate inside messy real-world systems instead of in a chat box.
But the shift doesn’t remove humans from the loop — it changes the loop. If the “AI coworker” story lands, many knowledge workers will spend less time producing first drafts and more time supervising, auditing, and coordinating. The next productivity battleground won’t be who has the smartest chatbot. It will be who can make agents reliable, governable, and genuinely useful inside real organizations.
Sources
Ars Technica:
https://arstechnica.com/information-technology/2026/02/ai-companies-want-you-to-stop-chatting-with-bots-and-start-managing-them/
Anthropic announcement (Claude Opus 4.6):
https://www.anthropic.com/news/claude-opus-4-6
Claude Code docs (Agent teams):
https://code.claude.com/docs/en/agent-teams
OpenAI Frontier:
https://openai.com/index/introducing-openai-frontier/
←
Previous Post
Next Post
→
oEmbed (JSON)
oEmbed (XML)
JSON
View all posts by Admin
Bing delisted 1.5 million Neocities sites. Here’s why that matters for the open web
The “Virtual Console” is back… sort of: what Nintendo’s Switch 2 retro downloads really mean
AI vendors are shifting from single chatbots to teams of tool-using agents. Here’s what agent teams are, why they’re emerging now, and why supervision and governance matter.
Document Title
Page not found - Florin.blog
Image Alt
Florin.blog
Title Attribute
Florin.blog » Feed
RSD
Skip to content
Placeholder Attribute
Search...
Page Content
Page not found - Florin.blog
Skip to content
Home
Blog
Garden Decor
Indoor
Main Menu
This page doesn't seem to exist.
It looks like the link pointing here was faulty. Maybe try searching?
Search for:
Search
Quick Links
Outdoors
About
Contact
Explore
Bestsellers
Hot deals
Best of The Year
Featured
Gift Cards
Help
Privacy Policy
Disclaimer
: As an Amazon Associate, we earn from qualifying purchases — at no extra cost to you.
Florin.blog
Florin.blog » Feed
RSD
Search...
l Slovenščina