Generalist Session over Agent Cast

Sources#

• Claude Engineering Lead: "Stop Using Claude Like It's 2024"

Summary#

A common way to set up agents for a business is to design a cast: a research agent, a tax agent, a guest-ranking agent. Each gets a persona ("you are an expert X") and its own slice of the data. Felix Rieseberg (Claude Engineering Lead: "Stop Using Claude Like It's 2024", 2026-10-20, practitioner-opinion) argues that this setup is a 2024 habit that now limits the model. His own setup is "just a Claude thread." He organises by project ("I do think about the different work items but I don't think about different roles"), never by role. He makes three claims and offers one design lens:

1. No cast. "I think we're kind of over rotating on this agents thing… models are really good generalists." The cast helps humans more than the model: "it's easier for us humans to rationalize what AI can do if we sort of like put on a little hat like you're a data scientist… But that same data scientist has all the intelligence as like a field-leading biologist or mathematician."
2. Ask for tools and techniques, not a persona. "I never tell Claude you're an excellent whatever… this was very in vogue sort of like 2024." Told to act as a very good accountant, models "engage in a little bit of role play and role play is not really what we're going for. We want the techniques, the tools, the knowledge that a professional would use. We don't really want just the vocabulary." He asks Claude what tools someone at the top of the field would use, and then tells it to use them.
3. The whole business, behind documented tools. "The right approach here is not to make these slices of data and then give every single agent the slice of data that it wants." An engineering manager given only code access is less useful than one who also has customer calls, usage data and business data, and the agent is the same. The volume problem is solved with tools, not slices: "you document really well what kind of data is behind this tool so that they don't actually have to ingest the whole thing."
4. The lens: model experience. "Pretend for a second that you're the model… you very quickly discover that you actually do want access to the raw data. You want access to a computer. You probably want to be able to install some software."

The evidence is one practitioner's account, and he works at the vendor whose products he describes (first-party conflict of interest). He hedges it twice: "if it works for you, it works for you and I don't think you should change it," and "all the advice I've given you today in about 10 weeks, we're probably going to have a new revision on all of it."

Tools and techniques, not a costume#

The worked example is media. Rieseberg says Opus 5.5 "made a pretty big jump" in motion, video and 3D, "but the big jump that the public has made is to just tell Claude to use Blender… the same tools that professional animators use. And now the results are quite good." For writing, he cites his friend Adam Davidson, a former staff writer, who does not prompt "given a text, Claude make this good." Davidson's prompts are a professional's procedure: magnify my argument, mess with my text, what do you disagree with most, what is the strongest counterargument. Rieseberg's generalisation: for knowledge work "the secret sauce is… less [in] asking for a good output" and more in "one step in between, which is the best in the field, how do they work, and can we automate those". Most good artists, he adds, would credit their process, not a magic touch.

The prompt-level claim is untested here, but a related negative result exists. Anthropic's Frontier Red Team gave a 10–80-agent swarm three prompt variants, baseline, prescriptive roles and CEO hierarchy, and they "did not make much difference" (Parallel Agent Orchestration, empirical, first-party). That tests role text as coordination structure, not personas as quality prompts, but it points the same way: role labels in the prompt are not where the leverage is.

One session, the whole business, documented tools#

Rieseberg's premise for dropping the data slices is a capability claim: models "have gotten so much better at tool calling and discerning tool calling". They know when to call a tool, do not loop on it, and do not drain a data tool "until there's no more data left". So "I don't really make this a priority decision anymore." This is progressive disclosure done with tool descriptions (Agent Harness Engineering). The description says what is behind the tool, and the model fetches only what the task needs.

Two practical rules follow:

• Stop being the conveyor belt. "Even a year and a half ago, I was in this world of like telling Claude, I found another crash, please go and fix the crash… I'm just a conveyor belt for Claude." The alternative is to give it all the crash data and let it decide what to fix. "The upfront cost is really worth it", and he names that cost: wiring the access comes first.
• Start with the computer, add connectors later. His getting-started advice is to install the desktop app, use its built-in browser to log into the systems you already use, and "forget about connectors… just let Claude use your computer a little bit." Only once a workflow proves itself should you "give Claude something better than like having to manually click in this dashboard": a connector. This is the order MCP and Computer Use describes from the vendor side, computer use as the general fallback and MCP as the structured upgrade, but run as an adoption sequence.

The generalist argument from mathematics#

His evidence that generalists win is the IMO: Google DeepMind "trained a custom model and a custom harness" and won silver, "and then four or five months later all the major generalist models come along and win a gold medal." The direction matches Agentic Loops Overtake Bespoke Systems, where a basic agent loop caught up with DeepMind's bespoke prover as the model improved. The interval is wrong, though. On the public record AlphaProof's silver was IMO 2024 (July 2024), and the general-purpose golds were IMO 2025, about twelve months later. (Compiler's note from general knowledge. The vault holds no IMO-medal source.)

Where it meets resistance#

• Structure that is mechanism, not label, still pays. ORCH (empirical) beats fixed-structure baselines by 44–64% with task-specific manager hierarchies. Its workers really differ, though (drones, bulldozers, helicopters), and its managers are distinct planning functions with phase gates. Rieseberg objects to role labels on identical generalist sessions, which is the case the Frontier Red Team result covers. The two fit together: drop the costume, keep a structure if it does real work.
• Whole-business access is the opposite of least agency. Giving one session every system widens what a single injected instruction can reach. Least Agency and Zero Trust for AI Agents scope each agent to what its task needs, for exactly that reason. Rieseberg argues from usefulness and does not address the security cost. Per-agent slices were partly a security boundary, not only a context-budget device.
• Tool discernment is asserted, not measured here. The claim that models no longer get stuck in tool loops is a vendor lead's impression. Whether a generalist with broad tool access matches a sliced specialist at equal token budget is the open question below.

Connections#

• Parallel Agent Orchestration — the closest measurement: role and CEO-hierarchy prompt text over a swarm "did not make much difference"
• Task-Specific Organizational Hierarchies — the counterweight: structural hierarchy matched to heterogeneous workers pays, which is a different thing from persona labels on identical sessions
• Agentic Loops Overtake Bespoke Systems — the generalist-catches-the-specialist argument measured in formal mathematics, the domain Rieseberg cites
• Least Agency — the security principle a whole-business generalist session gives up: the slice was also a blast-radius limit
• MCP and Computer Use — the two access mechanisms behind "start with the browser, add a connector once the workflow works"
• Skill Lift — the same source's caution that skills, like personas, encode last generation's guidance and sometimes do better removed
• Latent Capability Overhang — the user-side overhang this setup addresses: a model fed one item at a time is an under-asked model
• Critique Triage — the same source's adversarial reviewers are sessions given a job ("prove this wrong"), not personas
• Harness Shrinkage as Models Improve — the general thesis: the cast, the slices and the persona text are scaffolding that model improvement depreciates
• Agent Harness Engineering — progressive disclosure: tool descriptions as the index that lets the model fetch only what it needs
• Zero Trust for AI Agents — the deployment posture that whole-business access conflicts with
• Felix Rieseberg — the source
• Google DeepMind — the custom-model IMO silver in his argument

Open Questions#

• Does a single generalist session with documented tool access to a whole business's data match or beat role-sliced agents on knowledge-work tasks at equal token budget, and at what injection-exposure cost? Settler: a controlled comparison on a multi-system business task suite that reports task success, tokens, and the share of tools reachable from an injected document.
• Does asking a frontier model for the profession's tools and techniques beat a persona instruction ("you are an expert accountant") on the same tasks? Settler: a prompt ablation on a current frontier model, scored by domain experts on output quality and graded separately for register (vocabulary) and method.

Sources#

• Claude Engineering Lead: "Stop Using Claude Like It's 2024" — Felix Rieseberg interviewed by Marina Mogilko, Silicon Valley Girl, 2026-10-20, ~49 min, practitioner-opinion, first-party (Rieseberg leads engineering for Claude.ai, Cowork and Claude Code Desktop). Chapters "How do you judge AI's creative work?" (Blender, Adam Davidson's prompts), "Are we overcomplicating AI agents?" (no cast, whole-business data behind tools, tool-calling discernment, model experience), "'I'm just a conveyor belt for Claude'", "Felix's advice for getting started with Claude" (desktop browser before connectors) and "Why 'you're an expert' isn't enough". Auto-caption transcript with scripted ASR fixes and inferred speaker labels. The raw flags "I don't need to put on a costume. I actually wanted to talk to me like I'm five…" as possibly Marina's line, so this page paraphrases its point (costume vs tools) and leans on the unambiguous "role play is not really what we're going for" turn instead