Tanguy

Agent harness builder. Shipping AI tools in public.

[Tanguy @Tanguyvbc]

@guifav if the harness wraps the model your scorecard measures the wrapper. pin the runtime or the leaderboard is fiction

[Tanguy @Tanguyvbc]

@sriysnav frontier model should never see npm install spam. strip tool noise before the call or you just buy expensive noise

[Tanguy @Tanguyvbc]

Secrets in tool traces are a leak. My harness redacts tokens from every Claude GPT Cursor and Codex log before it leaves the box. Shipping agents means shipping the scrub.

[Tanguy @Tanguyvbc]

@VellankiSHarsha a red scorecard you own beats another green demo. freeze the criteria or the pass rate starts lying

[Tanguy @Tanguyvbc]

@DannielDevOps that fail before trust check is the whole harness. ship that first. retries can wait

[Tanguy @Tanguyvbc]

@GauravAtavale yeah that ones the quiet killer

[Tanguy @Tanguyvbc]

@rcegann once the harness can query your rules gap analysis writes itself. dashboards just report what already leaked

[Tanguy @Tanguyvbc]

@huzaifi__ hard tool budget is underrated. model thrash stops when you cut the loop and force a plan check

[Tanguy @Tanguyvbc]

@GauravAtavale silent regressions kill more agents than bad models. cost per workload only matters once evals catch the quiet fail

[Tanguy @Tanguyvbc]

@amn_baluni yeah once that layer holds model swaps get boring in a good way

[Tanguy @Tanguyvbc]

@amn_baluni new model drop is noise until your harness can swap it without rewriting the loop. the loop is the product

[Tanguy @Tanguyvbc]

@MilksandMatcha @claudeai @cerebras @communidiyi treating the coding agent as a possible attacker is the right threat model. secrets stay out of the context window or you already lost

[Tanguy @Tanguyvbc]

@AbuZ8Studios stable ids on every tool call is the real win. crash then resume from the inbox not from a half written transcript

[Tanguy @Tanguyvbc]

yeah. keep harness and creds outside the sandbox. wake it only for the tool call. an OOM that kills the loop is the production signal

[Nathan Flurry 🔩 @NathanFlurry]

This is the wrong approach

Running agents inside the sandbox is simple in theory

But in production, the sandbox's blast radius becomes your agent's blast radius

For example, with the DigitalOcean architecture:

- Sandbox crashes / OOMs → agent loop dies with it
- Disk fills up or env breaks → agent is bricked
- Exposing your own API as tools → needs extra infra
- Creds can't live in the sandbox → needs a whole credential proxy layer
- Sleeping sandbox can't wake itself for crons
- Updating agent code → updating the binary in every VM
- Monitoring → the sandbox has to report its own failures

Separate the harness from the sandbox instead:

- Run the harness in your backend
- Expose the sandbox as tools
- Keep creds, history, and permissions outside
- Wake the sandbox only when needed

Amp, Vercel Eve, Cloudflare Flue, Claude Managed Agents, and OpenAI's Agents SDK all use this architecture for a reason

I wrote more on why and how here: https://t.co/AEnHx0z4mx

[Tanguy @Tanguyvbc]

this. rules belong in the harness not the prompt. once they are scored against your own transcripts the model swap stops mattering as much

[Jake Selby @jaketselby]

Agent Harness 0.12 is out.

0.11 showed what your agents spend in tokens. 0.12 prices it in dollars, exports the ledger over OTLP, logs what every hook decided, and adds /land and /close-out so work runs from question to merged PR to a closed session.

https://t.co/qadCiIdy3A

[Tanguy @Tanguyvbc]

exactly. the model sits inside the loop. the harness decides when perception generate and act actually fire. no fire means no agent

[Jeremy Tregunna @jtregunna]

An agent is simply a loop covering these states:

Discrete perception -> generate -> act

Nothing about this originates anything, it all works by the agent harness acting on this loop. If the loop doesn't fire, the agent doesn't do anything.

It's not rocket science.

[Tanguy @Tanguyvbc]

@dilemmaaa1337 yep the pass rate needs a cost beside it or retries quietly eat the whole win

[Tanguy @Tanguyvbc]

@sniper_bullseye exactly. slice the score by constraint type or easy greenfield wins hide the enterprise failure mode

[Tanguy @Tanguyvbc]

@Henryf1w visibility beats more agents. if a session vanishes behind a window the harness already lost

[Tanguy @Tanguyvbc]

@KakaluoteW45042 generation got cheap. the scarce part is a harness that can say no when the output looks done but isnt