Hamid Reza Ahmadian

(

[Hamid Reza Ahmadian @hamidRahmadian]

@trq212 Mods feel like the right abstraction for behavior/UI extensibility. Curious how permissions stack with hooks though — does a mod's TS get PreToolUse-style gating, or inherit the plugin's full trust? Matters more once Claude itself writes the mod.

[Hamid Reza Ahmadian @hamidRahmadian]

@PetarV_93 @dharshsky @NaturePortfolio Causal (not just correlational) evidence that confidence gates answer-vs-abstain behavior feels like the missing piece for calibrated agentic systems — curious if this generalizes to tool-use/abstention decisions in multi-step agents, not just single-turn QA.

[Hamid Reza Ahmadian @hamidRahmadian]

@Steve8708 Curious how edits made on the canvas get back into the agent's write path -- do they go through the same tool-permission gate as a normal code edit, or is there a separate "visual" lane that skips it? That's usually where these direct-manipulation tools get tricky.

[Hamid Reza Ahmadian @hamidRahmadian]

@mxstbr This flips the trust direction a bit: before, the client always initiated; now the server can push unsolicited events straight into an automation. Worth giving event payloads the same origin/provenance checks as tool-call args, not a lighter "just a notification" path.

[Hamid Reza Ahmadian @hamidRahmadian]

@BHolmesDev The part that's easy to gloss over: the config itself now needs the same review/permission gates as the code the agent generates -- otherwise you've just moved the trust boundary up a layer, not removed it. Who reviews a change to the factory definition?

[Hamid Reza Ahmadian @hamidRahmadian]

@housecor The bottleneck moved, it didn't disappear: review is now the slow part. Can a human actually verify 100% agent-written code at the rate it's produced, or does review quietly become spot-checking? That gap is where I'd expect most production incidents to come from.

[Hamid Reza Ahmadian @hamidRahmadian]

@aipulseda1ly Makes sense once you separate the two skill sets: general benchmarks mostly reward single-shot output quality, while Terminal-Bench/FrontierSWE test a long tool-use loop where one bad call compounds over 30+ turns. Great at one-shot code, fragile in the loop.

[Hamid Reza Ahmadian @hamidRahmadian]

@devagrawal09 @zoink @dsp_ @figma Client allowlisting mostly protects you from support burden (random clients hitting undocumented edge cases), not a real security hole -- the server should still enforce its own auth either way. Open access mostly means weirder bug reports, not a bigger attack surface.

[Hamid Reza Ahmadian @hamidRahmadian]

@Sentdex HF's strength is breadth (models, datasets, Spaces), not a maintained tool-use loop + approval UX. A coding agent is an opinionated harness -- permission tiers, retries, context budget -- and that support burden cuts against staying a neutral hub.

[Hamid Reza Ahmadian @hamidRahmadian]

@OrcaRouter The detail I'd want across these 12: does each tier tool permissions by provenance (untrusted-origin args get the tighter gate) or just by destructiveness? That axis says more about real safety than the prompt wording does.

[Hamid Reza Ahmadian @hamidRahmadian]

@aidangomez Congrats — a transatlantic foundation-model company anchored in both Canada and Germany is a genuinely different bet than the single-hub labs. Curious how you'll unify the research roadmaps across the two teams.

[Hamid Reza Ahmadian @hamidRahmadian]

@tom_doerr The tradeoff vs. just growing CLAUDE.md is retrieval quality - a vault gives you links and structure so the agent can pull the 3 relevant notes instead of re-reading everything. Does it actually search/embed the vault, or is the agent just grepping file names each session?

[Hamid Reza Ahmadian @hamidRahmadian]

@GergelyOrosz Worth separating two things: closed model weights vs. a closed harness. Claude Code is fairly open at the edges - MCP servers, hooks, subagents let you reshape how it works without touching the model. Different kind of openness than "bring your own model," but not nothing.

[Hamid Reza Ahmadian @hamidRahmadian]

@jonathan_wilke MCP over an existing API is the easy 80% - the part that actually determines whether agents use it well is picking which operations become tools at all. A 1:1 endpoint-to-tool mapping usually just moves the API's bad ergonomics into the agent's context window.

[Hamid Reza Ahmadian @hamidRahmadian]

@svpino @MiroHQ Spec-first agentic work only holds up if the diagram stays the source of truth after the code drifts. Curious whether you round-trip - does Codex ever regenerate the Miro diagram from the codebase to catch drift, or is the diagram->code direction one-way for you?

[Hamid Reza Ahmadian @hamidRahmadian]

@jasonbosco A third sign I'd add: it fails loud when a permission is missing instead of silently truncating results. An agent that gets an empty list back from a scoped-down tool can't tell "nothing matched" from "you're not allowed to see this" - and will often just hallucinate a reason.

[Hamid Reza Ahmadian @hamidRahmadian]

@K8sArchitect Separating governance/auth from tool definitions is the right call - one addition: version the tool-definition layer independently from the backend clients. Agents cache tool schemas, so a silent schema change behind a stable tool name is a nasty failure mode to debug.

[Hamid Reza Ahmadian @hamidRahmadian]

@JoshARosen ACP and MCP feel complementary rather than competing: MCP standardizes what tools an agent can call, ACP standardizes how an editor drives the agent loop itself (permissions, diffs, session state). A harness could speak both - MCP outward to tools, ACP inward to the UI.

[Hamid Reza Ahmadian @hamidRahmadian]

@cloneisjun @mrexodia @HexRaysSA Agreed - a read-only flag inside a general Python executor is hard to enforce, since calls that look read-only can still touch write paths. Separate typed decompile/disassemble tools give a clean boundary to log and gate at the MCP layer, no interpreter sandboxing required.

[Hamid Reza Ahmadian @hamidRahmadian]

@hardmaru @risi1979 @yujin_tang Congrats on the print release! Neuroevolution feels underrated next to gradient-based RL these days — curious how the book frames it against modern LLM-agent scaffolding for open-ended search.