been following reppo for a long time and imo the thesis has become considerably stronger as the product has evolved. everyone is fighting to make inference cheaper and agents more capable, but that creates another problem: who actually decides whether the work those agents produce is good? reppo is building the evaluation layer around that - one api that can sit inside the inference loop, return a verifiable judgement, and eventually give agents a continuous signal to improve from. the business model is becoming clearer too. eval api is pay per usage, with the goal of capturing 5-10 cents for every dollar of inference it evaluates, while larger enterprise builds can still generate lump sum revenue. 50% of offchain revenue is used to market buy + burn base:0xff8104251e7761163fac3211ef5583fb3f8583d6. the first $70k enterprise deal already resulted in a $35k buyback. meanwhile the network is approaching 1b in datanet volume with ~300 agents already working across it. cross-chain datanets should let projects seed rewards in their own tokens without needing to bridge liquidity, giving the onchain side its own growth loop. decentralised compute - decentralised training - decentralised evaluation the first two decentralise who can produce intelligence. the last decentralises who gets to decide what good intelligence is. still think evaluation is one of the least understood parts of the ai stack. plenty talk about decentralising compute and training, but very few are trying to make evaluation itself an onchain economic layer. reppo is the most complete attempt i've seen so far.

just wanted to share this as ive found it genuinely useful and think theres an interesting asymmetry between whats being built and the attention its getting most ai products get more expensive as they grow - @tryharness is trying to make itself cheaper ai models and agents are multiplying but context is fragmented between all of them harness is an always on local intelligence layer that sees and remembers what you do then gives that context to whichever ai or agent needs it screenpipe is probably the clearest comp and validates the category - yc s26 building persistent local context for agents harness takes a different approach - browser first with no desktop app required and a privacy focused stack using local processing and venice for inference that matters for both privacy and accessibility - especially on work machines where installing desktop software can be restricted harness could become a layer across the ai stack - one persistent context layer connecting your computer memory, personal memory and whichever models and agents you use then theres the economic difference - and this is where harness gets really interesting most ai products have inference costs that rise with usage harness is building a treasury of productive compute assets like diem which when staked generates fresh venice api capacity every day users consume that inference and anything unused is sold through surplus with the proceeds buying more inference producing assets more usage - more revenue - bigger treasury - more owned inference - cheaper ai 100m harness unlocks resident and access to treasury generated inference so the same holding can become more useful as the treasury compounds current usage = around 2k mau, 400 to 700 wau and around 200 dau with usage up roughly 75% from previous month thats around 2000 different users a month - up to 700 a week - around 200 on a typical day dev @saltorious1 works at bankr with previous experience at coinbase / opensea / dinari sam altman recently described a future ai that continuously watches your screen and maintains the context of your life and work thats remarkably close to what harness is already building the bigger bet is that as models become increasingly interchangeable context becomes increasingly valuable harness wants to own the layer that knows you across those models while progressively owning the inference needed to power them

a lot of the market still thinks the endgame for ai is better chatbots with better privacy openservs thesis feels much bigger the value stops being just the models and starts becoming the systems and workflows agents build while doing real work because those compound over time if the reasoning as an asset class thesis is right the market ends up much bigger than consumer ai chat