Dr Fazal Ali

No one can shield their AI chats from pros­e­cu­tors pur­su­ing charges. Con­ver­sa­tions with AI as­sem­blages could be de­mand­ed by ad­ver­saries in lit­i­ga­tion, in both crim­i­nal and civ­il cas­es. Chats with ad­vo­cates are al­ways con­fi­den­tial. But AI as­sem­blages are not lawyers. Con­ver­sa­tions with AI as­sem­blages can erase the le­gal pro­tec­tions known as the at­tor­ney-client priv­i­lege, which usu­al­ly shield con­ver­sa­tions be­tween clients and their ad­vo­cates.

It is ill-ad­vised to use AI as­sem­blages to pre­pare state­ments, as such in­ter­ac­tions may in­clude de­tails from lawyers shap­ing the de­fence. Claude ex­press­ly ad­vis­es users that they should have no ex­pec­ta­tion of pri­va­cy in their in­puts. For­ward-lean­ing Cham­bers are ad­vis­ing their clients to se­lect AI plat­forms care­ful­ly, and they pro­vide prompt en­gi­neer­ing guardrails. Dig­i­tal life de­mands a deep ap­pre­ci­a­tion of the dif­fer­ent ecolo­gies and ar­chi­tec­tures of all AI of­fer­ings. Each one is plac­ing a dif­fer­ent bet on the fu­ture. Google, Ope­nAI, and An­throp­ic of­fer ca­pa­ble mod­els. Claude, GPT, and Gem­i­ni can write code, analyse doc­u­ments, and solve mul­ti-step prob­lems us­ing prompt en­gi­neer­ing. But how they ap­proach AI Agents sets these AI providers apart.

The fu­ture will be mul­ti-mod­el, mul­ti-in­fra­struc­ture, and mul­ti-agent, with Google, Ope­nAI, and An­throp­ic each hav­ing a dif­fer­ent vi­sion for AI Agents. Google is bet­ting on da­ta ac­cess and plat­form depth. Ope­nAI is bet­ting on ver­ti­cal in­te­gra­tion, and An­throp­ic is bet­ting on safe­ty as in­fra­struc­ture.

These dif­fer­ences are not mar­ket seg­men­ta­tion choic­es or brand iden­ti­ty dif­fer­ences. They re­flect what the AI ecolo­gies of each plat­form look like for de­vel­op­ers and builders, the trade-offs users are will­ing to ac­cept when choos­ing among them, and how these choic­es will evolve. An AI agent is not just a mod­el with mem­o­ry. The tool you choose will take ac­tion in the world. This means that tool use, plan­ning, fail­ure, mis­takes, re­cov­ery, and the han­dling of re­al con­se­quences are all con­sid­er­a­tions users must al­ways keep in mind.

For these rea­sons, AI be­he­moths are fore­cast­ing where their agents will work. This is the dis­tri­b­u­tion-and-de­ploy­ment di­men­sion of the ecol­o­gy of AI agent-mak­ing. The mod­el’s rea­son­ing ca­pa­bil­i­ties shape how it han­dles mul­ti-step plan­ning, blurred bound­aries, and un­cer­tain­ty. The ecosys­tem ar­chi­tec­ture con­cerns how Agents con­nect to tools, da­ta, and oth­er sys­tems. The au­ton­o­my phi­los­o­phy shapes how much free­dom the mod­el can be giv­en be­fore a hu­man is brought in­to the loop.

An­throp­ic has pub­lished clear guid­ance on how Claude should be­have as an agent. The core prin­ci­ple is min­i­mal foot­print. Agents must re­quest on­ly the per­mis­sions they need, choose re­versible ac­tions over ir­re­versible ones, err on the side of do­ing less, and con­form when the scope is un­cer­tain or vague. Agen­tic mod­els that delete files, send emails, and make API calls with fi­nan­cial con­se­quences can in­flict tremen­dous harm when they mis­in­ter­pret prompts.

One of An­throp­ic’s most strate­gic moves in the agent space is its Mod­el Con­text Pro­to­col (MCP). This open stan­dard has been adopt­ed by a sig­nif­i­cant num­ber of de­vel­op­er tools, code ed­i­tors, pro­duc­tiv­i­ty plat­forms, and cod­ing en­vi­ron­ments. If it be­comes the stan­dard for how Agents work with tools, re­gard­less of the base mod­el, then An­throp­ic will have huge con­trol over the en­tire AI Agent ecol­o­gy.

The MCP de­fines how AI agents con­nect to da­ta sources and ex­ter­nal tools. It acts as a gen­er­al con­nec­tor for agen­tic sys­tems. An­throp­ic is gam­bling on a fu­ture de­fined by safe­ty-by-de­sign. This means that the agents that will win will be the ones that hu­mans will trust, as agents be­come more au­tonomous in the in­fos­phere.

Google has a com­pre­hen­sive pro­duc­tiv­i­ty plat­form and the world’s best search in­dex. Hal­lu­ci­na­tion is the biggest prob­lem in agen­tic AI. Google’s so­lu­tion is ground­ing. Re­al-time ground­ing is a huge ad­van­tage for agents that work with chang­ing in­for­ma­tion, such as AI-pow­ered price tags, news, and mar­ket da­ta in screen economies.

Oth­er AI mod­els need web search plu­g­ins and cus­tom re­trieval pipelines to achieve some­thing com­pa­ra­ble. Google has built it in. Con­text length is a con­straint for agents that may be used to process lengthy le­gal sub­mis­sions, en­tire code­bas­es, and widened, long con­ver­sa­tion his­to­ries. Google has tak­en this fron­tier fur­ther than any­one.

Google has al­so re­leased an Agent2Agent pro­to­col, an open stan­dard for agent-to-agent com­mu­ni­ca­tion. This ac­com­pa­nies its open-source Agent De­vel­op­ment Kit for build­ing mul­ti-agent sys­tems on Gem­i­ni. With mul­ti­modal rea­son­ing across text, im­ages, di­a­grams, video, au­dio, and code, Gem­i­ni’s ca­pa­bil­i­ties are ar­chi­tec­tural­ly deep, orig­i­nal, and not retro­fit­ted.

Ope­nAI em­pha­sis­es care­ful bound­aries. The aim is to own as much of the AI stack as pos­si­ble. Each AI-as­sem­blage is a dif­fer­ent ecol­o­gy, a dif­fer­ent ar­chi­tec­ture, and a dif­fer­ent bet on the fu­ture, not a dif­fer­ent road. But if you don’t know where you’re go­ing, then I guess any road will do.

But re­gard­less of the AI agents we choose, our prompts can and will be held against us in a court of law.

Dr Fazal Ali com­plet­ed his Mas­ter's in Phi­los­o­phy at the Uni­ver­si­ty of the West In­dies. He was a Com­mon­wealth Schol­ar who at­tend­ed the Uni­ver­si­ty of Cam­bridge, Hugh­es Hall; the Provost of the Uni­ver­si­ty of Trinidad and To­ba­go; the act­ing Pres­i­dent of UTT; and the Chair­man of the Teach­ing Ser­vice Com­mis­sion. He is the Pres­i­dent of NI­HERST and an ex­ter­nal ser­vices con­sul­tant with the IDB.