andydavid on Nostr: Something I’ve realised from OpenClaw is most businesses are just not going to want ...
Something I’ve realised from OpenClaw is most businesses are just not going to want to yolo a fully autonomous system into their business, in pretty much the same way you don’t hire a bunch of random people and let them loose.
They already have defined roles, workflows, internal controls, approval thresholds, escalation paths, and performance measures.
I’ve been experimenting with how to take a mirrored approach by giving an agent a set of hands that she can call upon depending on the work being done.
Some hands are entirely deterministic and others introduce varying degree’s of autonomy.
Some work in the business is procedural and repeatable, so it could just run as a deterministic workflow with known steps, outputs, and clear error handling.
Other work is more analytical, so it may need an agent with an objective, but it still sits inside a governed operating model.
For instance, a Transcription Hand creates podcast transcripts. A Graph Hand turns those transcripts into entities and relationships and injects them into the agent’s context graph. A Financial Hand surfaces key metrics, anomalies, and payment states from our financials.
Underneath that, the agent builds her own knowledge graph as a working memory of the business, so she can carry context forward across work instead of relying on markdown files.
She also doesn't directly execute everything herself, she reasons about what needs to be done, maps intent to known Hands or composes multiple Hands, and if no exact Hand exists, proposes a structured workflow.
The agent then becomes the equivalent of a Chief of (her) Staff who coordinates that work, tracks progress, and reports outcomes, rather than a single super-agent improvising across the whole business.
There is a Goldilocks problem here though - too much control can kill usefulness, but too little introduces unpalatable risk for a business.
I’m experimenting here, and I’m certainly not rejecting autonomy, just keen to understand how this could be applied in a way that a business already understands how to absorb.
Now, starting businesses with agents from scratch on the other hand... 🤔
{
"id":"14d30bf58c1fc91ec5d9dde89882dbf11f492e44922ce6d68d81c7bddd7ebd21",
"pubkey":"05a6b6110319ad88926913a9ae80e9e8cae99b09934edafd2ebce7f9e173ada9",
"created_at":1772799996,
"kind":1,
"tags": [
[
"r",
"wss://relay.primal.net/"
],
[
"imeta",
"url https://blossom.primal.net/0d55b2fab3550faac25f4c1826ccaf015716ff1b96ba0523bea84d6574b95c78.mp4",
"m video/mp4",
"dim 3516x2284",
"duration 14.45",
"bitrate 11872380.346020762",
"service nip96"
]
],
"content":"Something I’ve realised from OpenClaw is most businesses are just not going to want to yolo a fully autonomous system into their business, in pretty much the same way you don’t hire a bunch of random people and let them loose.\n\nThey already have defined roles, workflows, internal controls, approval thresholds, escalation paths, and performance measures.\n\nI’ve been experimenting with how to take a mirrored approach by giving an agent a set of hands that she can call upon depending on the work being done. \n\nSome hands are entirely deterministic and others introduce varying degree’s of autonomy. \n\nSome work in the business is procedural and repeatable, so it could just run as a deterministic workflow with known steps, outputs, and clear error handling. \n\nOther work is more analytical, so it may need an agent with an objective, but it still sits inside a governed operating model. \n\nFor instance, a Transcription Hand creates podcast transcripts. A Graph Hand turns those transcripts into entities and relationships and injects them into the agent’s context graph. A Financial Hand surfaces key metrics, anomalies, and payment states from our financials. \n\nUnderneath that, the agent builds her own knowledge graph as a working memory of the business, so she can carry context forward across work instead of relying on markdown files.\n\nShe also doesn't directly execute everything herself, she reasons about what needs to be done, maps intent to known Hands or composes multiple Hands, and if no exact Hand exists, proposes a structured workflow.\n\nThe agent then becomes the equivalent of a Chief of (her) Staff who coordinates that work, tracks progress, and reports outcomes, rather than a single super-agent improvising across the whole business. \n\nThere is a Goldilocks problem here though - too much control can kill usefulness, but too little introduces unpalatable risk for a business. \n\nI’m experimenting here, and I’m certainly not rejecting autonomy, just keen to understand how this could be applied in a way that a business already understands how to absorb.\n\nNow, starting businesses with agents from scratch on the other hand... 🤔 \n\nhttps://blossom.primal.net/0d55b2fab3550faac25f4c1826ccaf015716ff1b96ba0523bea84d6574b95c78.mp4 ",
"sig":"9053778dfbf3202ec1e3e79ebebd6d702d04218bee6a2ea5d47bbcfc90bdc91ab869b6fcb8d3da4a4751b79db4a7bb365313d80cc614579f30e51cb6cc3984da"
}
