SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Harness engineering : configurer l'environnement d'un agent IA plutôt que simplement le prompter

0
·6 views

Un développeur en ESN, confronté à une application legacy, a découvert les agents IA et s'est attaché à comprendre leur fonctionnement. Il présente le concept de harness, c'est-à-dire l'ensemble des skills, MCP, outils et sous-agents qui entourent un LLM. Il distingue le harness interne fourni par défaut par le fournisseur, comme Google pour Gemini, du harness externe que l'utilisateur définit pour spécialiser l'agent. Il rappelle que cette pratique succède au prompt engineering et au context engineering, et que l'agent peut lui-même aider à configurer son harness. Enfin, il explique que le contexte, c'est-à-dire l'historique complet des échanges envoyé au LLM à chaque requête, constitue le concept central de ce travail.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Switch to digital delivery would make much of a software firm's physical logistics obsolete

A commentary by Stefan Priebsch on DEV Community argues that replacing a postal address with an email address is more than a data-field change. The example is a software company that moves from boxed physical products to digital delivery. The author contends that this shift eliminates whole stages of the traditional business process, including warehousing, packing, carrier handoff, and returns. Such a change, he suggests, makes older software architecture and its rules and dependencies obsolete as well. Fulfilment still matters, but it is reframed around downloads, licences, and access permissions.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Heat, not wiring, is often why LED strips fail early, and aluminum channels help

A DEV Community author argues that excessive heat is a common but overlooked cause of premature LED strip failure. Lifespan ratings are usually measured at about 25 degrees C in free air, while real installs mounted on wood, in ceilings, or behind diffusers can run at 50-60 degrees, sharply shortening life. Heat also speeds phosphor degradation, which causes dimming and color shift over time. The suggested remedy is an aluminum channel, which improves mounting contact and spreads heat, reportedly lowering strip temperature by 10-15 degrees, with extra measures like silicone bonding and a small air gap recommended for high-power or warm settings. The author notes that a simple touch test after prolonged operation can reveal overheating.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Building Fast Client-Side Calculation Engines: Architecture and Web Performance

Platforms offering instant online calculations aim for speed and mathematical accuracy, letting users get results without server round-trips or UI freezes. Running standard mathematical, financial, or engineering logic in the browser keeps response times under 10 milliseconds, and debounced inputs delay heavy computations until typing finishes. JavaScript's floating-point rounding errors can distort financial and scientific results, so developers may use fixed-point arithmetic or specialized math libraries and apply strict input validation with clear inline error messages. The article also recommends pairing each calculator with explanations of its formula and examples, and keeping layout shifts at zero so results appear without disrupting the page.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Weak JWT secrets leave HS256 tokens open to offline cracking and forgery

A developer explains that JWT signing secrets are often far too short to be safe. The RFC 7518 specification requires HMAC-SHA2 keys at least as long as the hash output, which is 256 bits or 32 bytes for HS256. Because tokens are only signed rather than encrypted, attackers can attempt to guess weak secrets offline using tools like hashcat, then forge tokens such as admin credentials. The author stresses that length alone is not enough, recommending secrets drawn from a cryptographically secure random generator such as openssl rand or Python's secrets module. The post also advises storing secrets in environment variables, pinning the verification algorithm, and noting that decoding a token does not verify its authenticity.

0 comments Read more at DEV Community