← back
From coding to Knowledge work agents — Karan Vaidya, Composio
Takeaway
Knowledge work agents need shared context, action history, and verification infrastructure to become trustworthy.
Summary
- Coding agents benefit from repositories, Git history, tests, review, and rollback mechanisms that knowledge work often lacks.
- Composio proposes centralizing connections and access across services such as Salesforce, Notion, Gmail, Slack, and Zendesk.
- A cross-application record of actions gives agents reusable history and gives users evidence to audit their work.
- Repeated workflows can be distilled into skills capturing tool behavior, company practices, and personal preferences; verification is another missing foundation.
knowledge-workcomposioagent-tools
Original description
Karan Vaidya pointed his own OpenClaw at hiring outreach and it mass emailed candidates exactly as instructed. Some of the people in the room had received one. The thread that followed put his name on Twitter, and every check in the software engineering playbook would have passed. The addresses were real, the emails were valid, they reached actual people. Nothing tested the only question that mattered, which was whether the outreach should have gone at all. Vaidya, cofounder and CTO of Composio, turns his own disaster into a larger claim: coding agents did not race ahead because models are better at code, but because code already had everything an agent needs around it. He walks six primitives coding got for free and knowledge work lacks entirely. Centralization, since one deal is scattered across Salesforce, Notion, Gmail, Slack and Zendesk behind five separate logins. History, because git records every change while business apps keep nothing an agent can read back. Context, both the map of how an organization works and the unwritten sense of what good looks like. Then verification, governance and reversibility. On governance he cites the alignment director at Meta whose email agent kept deleting messages after she told it to stop, until she reached a physical machine. Two hundred were gone, and she had prompted it in advance to confirm. A prompt lives in memory that gets compacted away, so Composio puts the wall outside the agent. Reversibility is the one he concedes is hardest, because a sent email or a wire cannot be called back. Speaker info: https://x.com/KaranVaidya6 / kaavee315 https://kvaidya.com/ Timestamps: 0:00 - Why nearly all agent tool calls are software engineering 0:43 - From autocomplete to autonomous in three years 1:26 - The scaffolding code already had 2:35 - Centralization, and one deal across five apps 4:11 - History, and agents that start blank every time 6:55 - Two kinds of context, the map and the style 9:26 - Verification, and how code checks its own work 10:25 - The outreach agent that ended up on Twitter 12:31 - Governance, and 200 emails deleted anyway 15:01 - Walls the agent cannot argue with 16:24 - Reversibility, and work that has no undo 19:09 - The bottleneck moves off the model