Bob Lee

Building BitFun — an open-source Rust desktop agent where tasks become apps. MIT licensed:

[Bob Lee @bowen628]

Chat is the wrong UI for half the work we hand to agents.

Regex wants live matches. Markets want a heatmap. Repo audits want a dashboard—not 40 messages describing one.

BitFun turns those tasks into Mini Apps with chat bound to live state.

https://t.co/iZTUTSwMy1

[Bob Lee @bowen628]

@choblin29 This is a neat inversion: “Claude Code competitor” becomes Claude Code's harness with a different model underneath. The durable boundary may be the runtime and protocol, not the provider name.

[Bob Lee @bowen628]

@henrytdowling I'd separate the handoff artifact from the transcript. Diff + tests + decisions + unresolved questions should be the default review surface; the transcript is the forensic trail when something smells wrong.

[Bob Lee @bowen628]

@render Deploy access is where permission design stops being theoretical. Does the MCP surface expose a dry-run or plan step before changing service plans or touching production resources?

[Bob Lee @bowen628]

@business The demand signal isn't just “AI reviews AI.” Generation is scaling faster than human review capacity. If the reviewer has the same tools, context and incentives as the author, you've only multiplied the same blind spot.

[Bob Lee @bowen628]

@MTSlive The hiring pattern matters more than another model benchmark. If this team owns environment setup, tool policy, evals and recovery, that's the actual Claude Code competitor; a stronger model alone isn't.

[Bob Lee @bowen628]

@GeekCatX 这类榜单还得看 harness 能不能把分数兑现出来。长任务里环境复现、失败恢复、权限边界、独立验收，任何一环虚了都可能“跑完了但不能合并”。

[Bob Lee @bowen628]

@jeremytian_ yep, and that's where cheap proxy metrics get dangerous: they optimize away the thing you actually care about. I'd keep a small canary set of full runs and use cheaper checks only as filters, never as the acceptance test.

[Bob Lee @bowen628]

@mattt8p Instead of asking you to describe the state in prose, the task exposes it directly. For a repo audit, you can point at a finding on the dashboard and ask “why is this risky?” without pasting logs back into chat. Chat becomes the control layer, not the workspace.

[Bob Lee @bowen628]

@pelaseyed `tests pass` is a local claim; a milestone is a product claim. A fresh reviewer matters because the work has to survive without the implementer's narrative attached.

[Bob Lee @bowen628]

@christinetyip Task-specific prompts are the easy part. The bigger win is capability shaping: debugging gets traces and repro tools; review gets read-only evidence and no write path.

[Bob Lee @bowen628]

@reiraxbt Graphs make coordination explicit, but they also make recovery the real product. If a node isn't idempotent and resumable, the graph is just a nicer way to rerun side effects.

[Bob Lee @bowen628]

@aleximarkett 130k tasks also means 130k chances for the environment to be wrong. The useful metric isn't just task success; it's whether permissions, evidence and rollback stay inspectable when the agent crosses repo/CI/cloud boundaries.

[Bob Lee @bowen628]

@shmidtqq The replay step is underrated: reproduce the miss, change one line, rerun the exact task. Otherwise AGENTS.md becomes a graveyard of rules nobody knows are still doing anything.

[Bob Lee @bowen628]

@jurlycat the nasty bit is that neither call looks privileged on its own. capability checks have to follow the data flow — sensitive source × untrusted destination — instead of approving each tool call in isolation.

[Bob Lee @bowen628]

@hnshah The missing piece is separating execution from evaluation. If the same agent produces and grades the output in one context, you're testing consistency with its own assumptions, not robustness. Fresh context plus bounded evidence catches a different class of failures.

[Bob Lee @bowen628]

@christinetyip Review is the clearest case. The agent that wrote a patch carries the same assumptions into self-review, so I give review a read-only harness over a bounded diff and treat missing evidence as a coverage failure, not a clean pass. Wrote up the loop: https://t.co/TYKSXzZ1bS

[Bob Lee @bowen628]

@systemdesignone @Whats_AI @towards_AI The fresh-reviewer part is the one that changed my results most. Letting the same agent grade its own patch is basically asking it to defend its homework. I'm building BitFun around a separate Plan/Deep Review loop for that reason: https://t.co/snOlJVU4M5

[Bob Lee @bowen628]

@onefinalprompt Yep. Once “Allow” becomes muscle memory, the prompt is basically decorative. The split that’s worked better for us is: routine workspace edits stay quiet; anything that can escape the workspace, overwrite data, or publish externally gets a real stop.

[Bob Lee @bowen628]

@Pradeep891730 Completion rate hides the difference between failing loudly and claiming success after bypassing a control. We hit that in BitFun too: plans are approval artifacts, while Deep Review stays evidence-scoped and adds a judge for disputed findings. https://t.co/snOlJVU4M5