Most symmetric cryptographic algorithms requires a key that is indistinguishable from random. This means that the process used to generate the key must have a uniform, independent distribution over all bit-strings of the appropriate length. Using a non-uniformly-random key not only reduces the brute force needed to guess it, but may also open the door to related-key attacks.
The expressions “key agreement” and “key exchange” are often used to mean a mechanism like (elliptic curve) Diffie-Hellman, but (EC)DH itself does not produce a key, it only produces a shared secret. This secret is in some numerical range that depends on the curve: not all $2^n$-bit values are possible. Even within this range, it is not uniformly distributed (it can't be since there isn't the same number of private points and public coordinates).
To go from the shared secret to a key, you need to use a key derivation function. A KDF takes a secret which has a cryptographically high number of possible values but may not be uniformly distributed, and has output that is indistinguishable from random to anyone who doesn't know the secret.
In a pinch, if you only need one symmetric key from the key exchange, hashing the shared secret with a cryptographic hash such as SHA-256 or SHA-512 is fine. If you need more key material than one hash's length, or if you prefer to use a standard construction that may be more robust in case a partial weakness is discovered in a hash function, there are (too) many standardized key derivation functions, for example HKDF (robust and popular), the TLS 1.2 PRF (a variant of HKDF which is specific to TLS), the NIST SP 800-56C key derivation functions (some hash-based like HKDF, some AES-based), etc.